THE NVIDIA AI GPU BLACK MARKET | Investigating Smuggling, Corruption, & Governmentsjimmy_thang December 13, 2025 ## NVIDIA GPUs have become so in-demand for so-called "AI" workloads that a "black market" has emerged around them -- at least, in the eyes of the US Government. In China, it's simply a "market." We adventured on extensive travels throughout Asia and spent hundreds of hours investigating the issue The Highlights * We spoke to everyone about NVIDIA’s AI GPU black market, including middlemen who connect buyers and suppliers * The export of these GPUs to China is in violation of US Government law, which includes numerous restrictions on semiconductor processing capabilities * Fueling greed, manipulation, and propaganda, we think NVIDIA is playing all sides #### Table of Contents * AutoTOC Visit our Patreon page to contribute a few dollars toward this website's operation (or consider a direct donation or buying something from our GN Store!) Additionally, when you purchase through links to retailers on our site, we may earn a small affiliate commission. ### Intro We’re multiple administrations deep in a technological cold war over processing power between the United States and China. China’s Cyberspace Administration has labeled some US graphics processing products as a security risk, seeking answers about US government backdoors in the silicon. Meanwhile, the United States has imposed heavy restrictions on exports of graphics processing units, or GPUs, being sold to Chinese companies by American companies. The sale requires rarely-granted licenses for each import scenario to legally export GPUs above a certain performance level, with the stated objective being to restrict progress of private and government projects, including AI development, in China, while trying to maintain the US’ claimed AI leadership. The United States takes this so seriously that, just this week, the Department of Justice had two Chinese nationals arrested in California for what the DOJ alleges is the smuggling of tens of millions of dollars of GPUs. But where there’s prohibition, there’s smuggling. _Editor's note: This was originally published on August 17, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication. This particular story had a saga of what we felt was censorship behind it, thanks to Bloomberg L.P., which we've detailed extensively:_ _Part 1 (Our Channel Could be Deleted)_ _and Part 2 (YouTube's Systematic Punishment)_. _This was written to be seen as a video, so the adaptation sticks to a more viewable/colloquial use of language._ * * * #### Credits * * * ##### Host, Writing, Lead Editing Steve Burke ##### Editing Vitalii Makhnovets Tim Phetdara ##### Editing, Graphics Andrew Coleman ##### Camera Tannen Williams ##### Research and Writing Ben Benson * * * ### **Acquiring "Illegal" GPUs in China** _In China, it's not an illegal market -- it's just a market._ * * We spoke to everyone about this NVIDIA AI GPU black market: We found middlemen who connect buyers and suppliers, users who can understand the demand and explain it, using the most dystopian definition of wealth -- how many GPUs one has. We also found independent repair shops who, simply doing their jobs, salvage valuable silicon components from dead PCBs of banned GPUs, innovatively hand-modifying them to be better than and have more VRAM than NVIDIA’s own official product SKUs. These shops are not explicitly a part of any "black market," they're just repair shops that happen to sometimes work on export-controlled GPUs. We met multiple people who, when asked the same question, gave the same passcode-like Chinese idiomatic expression, or chengyu (成语), which translates to “open one eye, close one eye.” In other words, it means, “to turn a blind eye.” We heard this saying so much that we made a T-shirt based on it to help fund this investigative report. "There's a new kind of black market, and it's high-end AI GPUs. This particular black market is worth billions of dollars a year" And among others, we even spoke with a US-based Chinese national buying video cards to strip them and ship the GPUs to Chinese companies, which violates US export control law. One of our viewers was able to meet with a GPU smuggler, whom we’ll call “The Plug.” The smuggler spoke limited English, but they both understood one universal truth: money. He operated a GPU testing rig inside of a Prius that he drives around multiple states in the Western United States. The least suspicious thing in his car was a spare license plate in the trunk, but we’ll come back to “The Plug” towards the end. "Black market" is normally a phrase associated with drugs or guns, but there's a new kind of black market: high-end, AI GPUs. This particular black market is worth billions of dollars a year and is hiding in plain sight. Hong Kong is our first stop along our journey. By skyscraper count, Hong Kong would be the tallest city on earth. The density is unbelievable. We spent a few days here for this story, wandering markets and meeting sources. We went to Hong Kong to learn what the demand drivers are for the banned GPUs, how they get into China, and more about the illicit side of the GPU smuggling business. ### **Big Adventure  ** To provide a big overview of our big adventure, it began when we booked a 24-hour plane ticket to Hong Kong and 20 days of hotels across Hong Kong, Shenzhen, Zhengzhou, Huizhou, Taipei, and more. At least one of us got detained by at least one of the governments involved in this story, but we can't talk about these indecipherably singular or plural instances or instance of any aforementioned detainment. Then we talked to a lawyer about the previous sentence, then investigated whether or not GPUs are actually smuggled with lobsters into Hong Kong. * * We eventually found a pile of GPUs that are now export controlled and met a professor who really wants to make sure you know that they were “legally obtained.” After leaving Hong Kong, we booked a boat to Shekou, China. Then headed to Shenzhen, got kicked out of a warehouse, and used high-speed rail to go deep into China, where we met a guy who thinks that desoldering a GPU and reballing it is no big deal. We got a ton of information about GPU smuggling from a guy named “5.” After that, we went back to Shenzhen for the third time in two weeks, went back to Hong Kong, flew to Taiwan, before finally getting back on the plane to the US. ### **Perspective** > We shot over 12 hours of interviews We spoke to a lot of people on the record for this story and learned about the dystopian world of high-tech GPU smuggling. We spoke to people ranging from owner-operator trading companies to professors of economics building datacenters. We even tried to talk to the US Department of Commerce, but they didn’t reply, and every person on the chain from the US Department of State had out of office auto responders because we coincidentally emailed them the same week of an especially problematic story that enigmatically involved namedropping. We wanted to connect as many pieces of this puzzle together as we can today, and that’s why we shot over 12 hours of interviews that we just spent weeks cutting down. The point is finding people who know people, and each person in our lineup today led us to at least one other person in this video, eventually building the full pipeline of smuggler-to-user. Although this story isn’t about drugs, it is about a different thing that billionaire executives get a high from: AI. But before we get to any of those interviews, we need to establish the basics of this geopolitical mess. The story is complicated, so we'll start with defining these key facts: Why these GPUs are banned, the new 15% license (which only applies to two GPU models), who buying and selling is legal or illegal for, and then the timeline. > We have over 400 pages of research that went into finding sources and understanding the laws. ### **Why They’re Banned** AI GPUs have been in the mainstream news constantly. This story has been a complete mess to follow. It has spanned years and two US administrations. We have over 400 pages of research that went into finding sources and understanding the laws. > Reports warned of NVIDIA product use in nuclear weapons research Here’s why governments care about AI GPUs: NVIDIA functionally holds a GPU monopoly in our sector of the industry, which is building computers to play video games. That now feels relatively innocent by comparison to AI. The company leveraged decades of gaming domination to build a foundation for what it now focuses on, which is making the most powerful GPUs for AI in the world. Reports warned of NVIDIA product use in nuclear weapons research, facial recognition technology allegedly used in Russia to suppress dissent and growing concerns of AI facial recognition use in the US for similar deployments, alongside reports of use in international spying and in drone warfare. NVIDIA finds itself in the middle of all of this. Even though NVIDIA disputes selling to some of these entities -- for example, it says that it doesn’t sell GPUs to Russia -- the products still find their way there. NVIDIA is making money one way or the other. Someone is buying it, maybe from someone else who bought it from someone else, and it may be transacted through smuggling. But NVIDIA does end up selling the device ultimately to somebody. Regardless of who they sell to, NVIDIA plays a big part in this worldwide obsession of AI. And we think it’s playing all sides, but we’ll talk about that more at the end. Besides, when there’s a gold rush, it’s better to sell the pickaxe than swing it. Although the US doesn’t talk too much about its own use of AI, it spends a lot of time talking about China’s. The US restricts NVIDIA’s GPUs through export control rules that ban the sale of certain GPUs into China. The restriction is for the sale of GPUs by American companies or companies that want to transact business in America to companies that are in China or the Chinese government itself. Some examples of export-controlled GPUs include gaming GPUs like the RTX 5090 and 4090, which are useful in AI applications (mostly for their high VRAM capacity), and data center/AI GPUs like the A100, H100, H200, and B100, as well as the others shown in the image above. This list is constantly in flux. There are some new and incoming exceptions for the NVIDIA H20 specifically, which has faced Schrodinger’s GPU ban depending on whether CEO Jensen Huang had a one-million-dollar dinner at Mar-a-Lago with Donald Trump on a given week. #### **GPU Export Control Timeline** Now, we’ll get into the timeline of GPU export controls across multiple administrations. We have a separate article with the full, bulleted timeline here. ##### Obama Administration and Early AI Talk Even at the end of the Obama Administration in 2016, the US Government was just starting to talk about AI in relation to national security in a Wired interview. Then-President Barack Obama said: “Developing international norms, rules, protocols, verification mechanisms around cyber security generally and AI, in particular, is in its infancy. You got a lot of non-state actors who are the biggest players. Part of the problem is that identifying who's doing what is much more difficult. If you're building a bunch of ICBMs, we see them. If somebody's sitting at a keyboard, we don't. And so, we've begun this conversation. A lot of the conversation right now is not at the level of dealing with real sophisticated AI but has more to do with essentially states establishing norms about how they use their cyber capability. Who are you more afraid of: big brother and the state or the guy who's trying to empty out your bank account? Part of the reason that's so difficult is that if we're going to police this wild west, whether it's the internet, or AI, or any of these other areas, then by definition, the government's got to have capabilities. If it's got capabilities, then they're subject to abuse. And, at a time when there's been a lot of mistrust built up about government, that makes it difficult.” ##### First Trump Administration and Biden Administration Those were the early days. > the Biden Administration took major action in 2022 by restricting exports to China, Hong Kong, and Macau The first Trump Administration started a commission advising Congress on maintaining AI leadership, including simply banning the sale of some advanced semiconductor equipment and chips to China. Years of back-and-forth, a pandemic, an election, and Chat GPT’s launch later, then the Biden Administration took major action in 2022 by restricting exports to China, Hong Kong, and Macau. NVIDIA’s Ampere architecture A100 and newer codename “Hopper” H100 GPUs and systems were all restricted. NVIDIA shed tears for $400 million worth of lost sales as a result and was especially sad when many of its export-compliant alternatives to these also got banned, like its newly-created A800, H800, and L40S, in addition to NVIDIA’s RTX 4090 gaming card. NVIDIA said it didn’t expect “near-term meaningful impact” on its financials. ##### Second Trump Administration NVIDIA responded by designing a workaround to the previously worked-around workaround, leading to the H20. Then the government added new rules for high memory bandwidth cards and the Biden admin tried to come up with an AI chip diffusion rule that would limit the quantity of GPUs being sold into different countries rather than only by processing power metrics, because the government really didn’t know how the fuck to measure these things and NVIDIA, knowing more about GPUs, could tweak any dial it wanted to just barely be compliant. Then DeepSeek came out and everyone panicked, stocks plummeted, and the government scrutinized the role of NVIDIA GPUs in it. In February 2025, Fiscal Year 2025 results were posted and NVIDIA’s Singapore revenue skyrocketed to 18% of total revenue based on customer billing location despite shipments to Singapore being claimed to be less than 2% of Fiscal Year 2025 revenue, which caused people to say “wait a minute.” Unrelated: Several GPU smugglers were arrested in Singapore one day after the fiscal year report was posted, which caused people to say “that makes more sense.” > In May, 2025, Trump implemented wide-sweeping tariffs and rescinded the Biden chip diffusion rule that would have limited how many AI GPUs Jensen could sell to other countries AMD spawned out of nowhere to say it wrote-down $800MM of inventory due to export controls. NVIDIA one-upped it with a write-off of $5.5 billion. In May, 2025, Trump implemented wide-sweeping tariffs and rescinded the Biden chip diffusion rule that would have limited how many AI GPUs Jensen could sell to other countries, Jensen then said, “It’s just an incredible vision. I think this is going to be a transformative idea for the next century for us. These 2 initiatives are completely visionary and it’s going to be transformative for America.” The H20 GPU that was created to comply with rules was still okay, then Jensen had dinner with Trump at Mar-a-Lago for $1,000,000, then the H20 got banned -- he must have chosen a bad restaurant. In July, Jensen Huang met with Trump and was permitted to sell H20s again, Huang went to China, China said Jensen’s GPUs have tracking devices and backdoors, NVIDIA denied that, Jensen went to Washington, Trump then spoke very highly of Huang (and Lisa Su), and more importantly, he ragged on Intel’s CEO. Intel’s CEO then went to Washington, so Trump likes him now. Tim Cook materialized from the infernal plane to give Trump a 24-karat gold “gift,” and “gift” is in quotes because that’s not what that’s called. Then we get to this past week, when Trump asked NVIDIA and AMD to pay 20% to the US government for sale SPECIFICALLY of the H20 and AMD Instinct MI308 sales, not all GPUs as some erroneously reported. Cousins Jensen Huang and Lisa Su negotiated Trump down to 15%, and now they’re allowed to sell two specific cards that were originally created to comply with the laws before they changed and somehow everyone walks away a winner. Except now China doesn’t want them anyway. That about sums it up. Except one last thing that happened as we were writing this: The Department of Commerce doesn’t yet know the legality of the deal, with Tom’s Hardware highlighting legal expert arguments over Article I, Section 9 of the Constitution. If you want the full details with all of the in-between, make sure to check out our massive timeline article. ##### **A New Law** There was also recent news about a 15% revenue share between NVIDIA and AMD with the United States Government for sale of some AI GPUs. Trump stated, “This is an old chip that China already has and I deal with Jensen who is a great guy and NVIDIA. The chip that we’re talking about, the H20, is an old chip. China already has it in a different form, different name, but they have it. Or they have a combination of 2 will make up for it and even then some […] but the H20 is obsolete. You know, it’s one of those things, but it still has a market. So I said, listen, I want 20% if I'm going to prove this for you, for the country, for our country, for the US. I don't want it myself, you know, every time I say like 747, I want. Yeah, for the Air Force. So when I say I want 20, I want for the country. I only care about the country. I don't care about myself and he said ‘Would you make it 15?’ So we negotiated a little deal. So he's selling an essentially old chip that Huawei has a similar chip, a chip that does the same thing and I said 'good, if I'm going to give it to you' because they have a, you know, they have a stopper, what we call a stopper. Not allowed to do it. A restricted is really known as a restrictive covenant.” If you were to read only the headlines, you’d think that this applies to all GPUs and that the ban is over, and then you might also think that a market in China, or in the US’ eyes, an illegal market in China, would cease to be so illegal. That’s not the case. This new 15% revenue share would, if it’s legal (and they’re __ not sure yet), allow NVIDIA to sell specifically the NVIDIA H20 GPU to approved Chinese entities. They likely can’t be on the entity list. It would allow AMD to sell specifically the Instinct MI308 GPU to approved Chinese entities. The proposed license would not affect any other GPU that currently does not have a license. The government hasn’t made clear yet if NVIDIA’s partners would also be permitted to make these sales. That means that other banned GPUs, including the RTX 4090, RTX 5090, H100, B100, B200, and so forth, remain banned. The H20’s extremely high 96GB memory capacity would enable large models to fit in memory and actually run, especially with multiple GPUs in a single rack, even if it’s slower. That means companies can achieve performance targets by stacking GPUs which are lower clock and core count but higher capacity. As for the newer Blackwell architecture GPUs, Trump stated, “Now Jensen also has, Jensen’s a very brilliant guy, and Jensen also has a new chip, the Blackwell. Do you know what the Blackwell is? The Blackwell is super duper advanced.” Let’s not give NVIDIA any ideas on new GPU names. Trump added, “I wouldn’t make a deal with that. Although it’s possible I’d make a deal, a somewhat enhanced in a negative way Blackwell. In other words, take 30% to 50% off of it. But that’s the latest and the greatest in the world. Nobody has it. They won’t have it for 5 years. On the Blackwell, I think he’s coming to see me again about that. But that will be an unenhanced version of the big one.” For now, these “Super Duper” GPUs are not licensed for sale. In short: The H20 and MI308 were compliant with the US Government’s original rules, then the rules changed during design & production and they were banned, then AMD and NVIDIA collectively declared over $6B in financial impact as a result. Then the US Government said, ‘Wait a minute, we can help with that if you cut us in.’ It reminds us of the unpredictability that we highlighted in our tariffs documentary. The H20 most certainly is not “obsolete,” though. It’s still very desirable in China, and with a lot of them, they become particularly potent. #### **How the US Determines GPU Bans** Setting a threshold for banned hardware should be objective since it can be tested. This is a graph from the Department of Commerce that visualizes the original threshold at which a computing product became automatically banned for export to parts of the Middle-East and China without a granted license. There have been some changes since, but back when this was made, the “Total Processing Performance” score on the Y-axis was used to determine cards in need of a license. The government needed a metric to calculate against, so it created its own. Accelerators and video cards have a lot of metrics in their spec sheets, including memory capacity (which is critical and as simple as a pass/fail for certain training and AI uses), memory bandwidth, GPU clock speed, GPU SM or CU count, TPCs, Tensor Cores, ROPs that are sometimes randomly missing on NVIDIA devices, TFLOPS, TOPS, PFLOPS, GFLOPS, gigabits, power, and more. So then, banning a product could probably be based on some sort of benchmark rather than a random metric from a spec sheet, but the US government, illustrating what an absolute clusterfuck this situation was and now remains, decided to instead multiply one random metric from a specsheet against the bit length of the operation being executed. FLOPS, or Floating Point Operations Per Second, and TOPS, or Tera Operations Per Second, are calculated by the company making the spec sheet and aren’t a great measurement of actual performance. These numbers are based on both marketing and whether we’re talking about FP8, half-precision FP16, single-precision FP32, double-precision FP64, or Tensor performance, so the government said, “the rate of MacTOPS is to be calculated at its maximum value theoretically possible” and “the rate of MacTOPS is assumed to be the highest value the manufacturer claims in brochure[s] for the integrated circuit.” So it’s not based on a bunch of real-world benchmarks of applications or something useful. The government also references MacTOPS as the theoretical peak of TeraOPS in multiply-accumulate computations. The Biden administration used this “Total Processing Performance” (or TPP) score in 2023, with the government later adding a “Performance Density” metric dividing the TPP by the die area in square millimeters. In other words, the government didn’t want NVIDIA to be able to sell more of a lower-performance GPU to make-up for the loss of high-performing parts with multi-GPU solutions. There are a lot of reasons this doesn’t capture the full picture, like sparsity, APIs, differing methods to calculate FLOPS, and different performance for different applications, but the government needed a way to define a threshold, so this is what it made. The limit was a TPP score of 4800, exceeded by even the RTX 4090 when calculated using Tensor performance. Now, if this doesn’t mean anything to you, that’s OK, because it probably doesn’t mean anything to people signing the laws either. Or maybe that’s not OK, but you get the idea. Something like a higher memory capacity, lower FLOPS performance GPU or even series of GPUs like RTX 3060 12GB cards might be able to get the work done more effectively if it only needs memory. Memory wasn’t factored into the TPP calculation. That was the point of the H20, but then the absurdity of the situation expanded by introducing an opaque memory bandwidth requirement. The Register wrote, “Unlike with previous export controls, the Commerce Department's Bureau of Industry and Security (BIS) hasn't issued specific guidance on how much I/O or memory bandwidth is too much.” So it seems that, across now two administrations, the United States is creating formulas based upon numerical calculations and then, when NVIDIA and AMD tweak numbers to fit within that box, it is retconning those rules in a guess-and-check process. ### **Experts & Roles** We spoke to a lot of people in this story across different languages, which made it complicated. The discussions come from: Dr. Vinci Chow, Senior Lecturer in the Department of Economics at the Chinese University of Hong Kong, responsible for building his department’s machine learning servers and sourcing GPUs from middlemen suppliers. Dr. Zǐ Háo Fù, a Research Assistant Professor at the Chinese University of Hong Kong and Affiliated Lecturer at the University of Cambridge. Zihao specializes in both computer science and linguistics. * * Various retail workers at the computer markets in Hong Kong to get a ground-level understanding. Zhou, Creative Director of Product at video card manufacturer Yeston. Yeston is not involved in any "black market" of GPUs and mostly deals with AMD; however, the company gave us a tour of its GPU factories to better understand production processes. An anonymous seller who goes by the pseudonym of “SILVER,” based in Shenzhen Bao’an and manages a warehouse in Hong Kong that receives and processes smuggled GPUs. Vincent, a resourceful fence who lives next to the warehousing and markets filled with GPUs and accelerators. He buys and sells these devices in Shenzhen Huaqiangbei. His job is to know people. Vincent’s Cousin is the fixer’s fixer. If someone needs a component to fix a broken video card or to build a new one, they go to people like him to get the integrated circuits. “Mr. 5,” a Bilibili hardware reviewer with a specialized focus on thermal solutions. Like us, Mr. 5 has had run-ins with NVIDIA that have ended in a soured relationship over disputes regarding independent reviews and editorial independence. Brother Zhang, a renowned Bilibili uploader (basically a YouTuber in China) who runs a video card repair shop in Zhengzhou. In addition to repairs, he regularly gets large orders from customers asking him to build them custom, unofficial, higher VRAM capacity NVIDIA GPUs for large language model tasks. Brother Zhang is not a direct part of any "black market," as, again, his business operates legitimately within China as a repair shop. It does, however, come into contact with GPUs the USA considers to be "illegal" for sale into China. Companies in Singapore and Taiwan who act as intermediaries between NVIDIA, NVIDIA’s partners, and companies in China. The Singaporean and Taiwanese companies are able to bring banned GPUs and servers in and re-sell them to Chinese companies, skirting export controls. We are unable to disclose their identities as there would likely be punishment from multiple governments and NVIDIA. The “Plug,” a US-based Chinese citizen who drives around the country buying hardware from American end-users and resells it to companies in China and Hong Kong. And a special thanks to our translator Raymen Wu of the BLK SODA agency in Taiwan. Grab a GN Tear-Down Toolkit to support our AD-FREE reviews and IN-DEPTH testing while also getting a high-quality, **highly portable 10-piece toolkit** that was custom designed for use with video cards for repasting and water block installation. Includes a portable roll bag, hook hangers for pegboards, a storage compartment, and instructional GPU disassembly cards. ### **The Smuggling Pipeline** Here's the pipeline. If a GPU doesn’t fall off the back of a truck in China after it was assembled, or if it isn’t a "QC defect" that disappears from the scrap pile, it may instead be moved by “ants” to get to China. There’s a saying in Chinese that we heard a few times (mayi banjia - 蚂蚁搬家), which translates to “ants moving” that represents a linear pipeline. Each ant in the fireline serves a specific role. It’ll help to name those to keep everything straight. Here’s what we came up with: The Source has access to GPUs. This could be as innocent as you unknowingly selling your card on Facebook Marketplace to The Plug. The Plug is responsible for acquiring from the original source and re-selling the hardware to the China-based distributors. In between, there’s a Mule. Sometimes this is the Plug himself doing a trip home. We learned that overseas students also regularly return with what are feasibly defended as personal GPUs, that they bought at a retailer like Best Buy, that may then get resold for markup and profit. In either case, the Mule gets the GPU into the country either by shipping it without interception or by hand carrying it. Next is the Middleman, receiving the GPUs and often interfacing with or managing the warehouses that store the cards. The Middleman buys from multiple Plugs, including factories that get rid of rejects with fixable or unimportant QC defects, then sells those devices to more localized distributors. That’s when we get to the Fence, who buys and sells GPUs between middlemen and warehouses to end users in China. Then, we have who we’re calling the Fixer, except this time, it’s literal. The Fixer is an optional step that may involve soldering and modifying a GPU to improve it beyond its original specification to make it more marketable for domestic AI uses. They might also just fix QC defects from the factories. Finally, we have the User. This is self-explanatory: The user is the demand driver, and oftentimes, large enterprises want dozens or hundreds of GPUs or more, while smaller users like the university may just want individual units or small batches. With everyone’s role named, let’s continue. ### **Demand Drivers** Our journey began in Hong Kong. #### **HONG KONG: Dr. Vinci Chow, Chinese University of Hong Kong** Upon arriving in Hong Kong, one of our GPU dealer informants slipped us a price sheet for a mix of GPUs that are both export controlled and not. Some of the banned ones on the list include the 5090, 4090, 4090 (48GB), A6000, A100, and A100 (80GB). The person who provided the price sheet works with smugglers. The problem is, we don’t know what the going rate for these GPUs is. So we hopped in a Hong Kong cab and headed across Victoria Harbor and over to the Chinese University of Hong Kong to meet up with Dr. Vinci Chow. We found him through a Reuters story from a couple years ago. He works in the economics department at the university and he’s responsible for having built many of the servers and machine learning systems that are in use daily. "Upon meeting Chow, he showed us some A100 GPUs, which he emphasized were, 'legally obtained.'" He has a whole blog detailing his process, including the difficulty of sourcing components and the ease with fixing problems when so close to Shenzhen, such as having custom PCIe riser cables built and basically done the same day. Dr. Vinci Chow is the right person to start us off with this story and educate us on where to go next. We’re here to understand the user’s perspective because before there could be any market for it, there has to be demand. And that is what the university and organizations like it generate. > it's illegal to export to China without [a] permit but it's completely legal on our side, right? There's nothing that says that you cannot buy a high-end GPU. > > - Dr. Vinci Chow Upon meeting Chow, he showed us some A100 GPUs, which he emphasized were “legally obtained.” He elaborated that he obtained them around when ChatGPT was released and got them for “quite cheap.” He expressed that, during this time, even the A100, which became export-banned in 2022, “was actually quite cheap.” He shared that he got the A100s for “less than 10,000 US Dollars.” Following ChatGPT’s release, however, Chow said that the prices went “crazy.” Pointing out the oddity of the export ban, Chow stated, “It's interesting because, you know, it's the export ban, right? So it's illegal to export to China without [a] permit but it's completely legal on our side, right? There's nothing that says that you cannot buy a high-end GPU. So, from our end, as long as we follow all relevant procedures, there's absolutely nothing illegal about, for us over here, to buy these GPUs. So that makes it for a very interesting environment.” He added, “These universities have been purchasing GPUs. It’s completely legal on our side, right? Yet this is clearly not supposed to happen from the perspective of the US government.” Regarding pricing and availability of export-banned GPUs, the professor stated, “If you pay enough, supply is there. Maybe not if you want to build a super cluster, right? For research, most researchers are talking about 1 or 2 GPUs. And if they have the funding, then it’s possible to obtain that. It’s just everything is more expensive.” Having built Chinese University of Hong Kong’s cluster of GPUs during COVID, he has first-hand knowledge of how much banned GPUs cost to get in Hong Kong and estimated that an H200 would cost around “$30K.” When we looked at our GPU price sheet, the H200 was going for 213,000 HKD, which amounted to roughly $29,700 USD. This corroborates his estimate. When we asked Chow how he thinks the export-banned GPUs get into China, he stated, “So these GPUs are almost certainly moved one at a time, right? It’s very hard to get a full HGX system.” Later, Chow introduced us to his friend and associate Dr. Zǐ Háo Fù, who works in the university’s linguistics department and has a background in computer science. He uses the university’s mainline datacenter to train large-language models. He shared that they currently use A40 GPUs, gaming GPUs, and sometimes have an RTX 6000 available (NVIDIA has multiple generations of RTX 6000, including Ada Lovelace and Blackwell). Zǐ Háo elaborated that memory is the “most important part for researchers.” When we asked Chow (hypothetically) if they could get whatever GPUs they wanted with limitless money, he explained that if they had millions of dollars before the GPU ban, they could get them. He also stated that, “With the export ban, [university] departments are not necessarily willing to publicly list their computing capabilities.” ##### **Datacenter** Chow showed off the GPU cluster he constructed. He explained that a single 8-GPU system can put out around 4,000 watts. For reference, this would require 2 standard US residential circuits to support at 20A per circuit or, as Dr. Chow says, would be about “two hair dryers.” He also shared that you can’t plug the GPUs into a standard wall socket as that would immediately blow the fuse and that they had to install 2 three-phase power within their GPU cluster room. Chow shared that they have roughly 30 A100-class GPUs and a bunch of 3090s and 3060s. This amounts to roughly 50-60 GPUs. Discussing the logistics of how the US determines which GPUs should be banned, Chow stated, “If you go back like five years ago, everyone probably would think that like, yeah, FLOPs is a very important metric to consider. But turns out when it comes to loading large models, what we all care about is how much memory, how much VRAM you have, right? You simply cannot load a model if you don't have the VRAM. So now H20 becomes a very attractive option.” He added, “the H20 has drastically lower FLOPs, but then it actually has more memory than the original H100.” For reference, the H20 has 96GB of memory, whereas the H100 has 80GB. Chow added, “So it’s actually in some sense more attractive.” While the idea of black markets might convey back-alley deals, Chow says that the export-banned GPUs are simply shipped to customers. When we asked if the GPU bans were effective, Chow responded, “It is effective in preventing the building of a very big cluster. It’s just not possible to get 100,000 of the GPUs.” > He mentioned that the H200s were also assembled in China, which raises an interesting question of what prevents these GPUs from 'falling off the line.' Chow elaborated on that, stating,  'I don’t understand how it works at all. How is the ban even working?' When we asked them why they think the US government cares so much, Zǐ Háo stated, “I guess they just want to delay the speed of other countries of training the model, but their ban is very weird.” When we asked them if the ban seemed targeted at military uses or Chinese companies in general, Zǐ Háo said, “I guess it’s in general.” Chow chimed and said, “I don't think, from the US government's perspective, there's this distinction between like Chinese academia versus Chinese military versus Chinese commercial. I don't think they really consider there's a distinction.” When we asked to see if Huawei’s hardware seemed compelling, Zǐ Háo shared that he doesn’t know people who use Huawei hardware, but it may be a fallback option in the future if consumers in China can’t get NVIDIA GPUs. > All these high-end GPUs are still manufactured in China, but they're not allowed to be sold in China. > > - Dr. Vinci Chow Chow explained to us that the A100 GPUs he acquired were “assembled in China.” He mentioned that the H200s were also assembled in China, which raises an interesting question of what prevents these GPUs from “falling off the line.” Chow elaborated on that, stating,  “I don’t understand how it works at all. How is the ban even working?” He points out that the box for the banned GPUs even “clearly state that they are manufactured in China.” Chow shared a theory on how banned GPUs are able to be sold in China, “My guess is there must be spares. I don't know. Spare SXM modules, spare casings, and then somehow these spare parts just get assembled into a complete GPU and get sold.” > I would be surprised if they don’t, right? I would be surprised. I would be really surprised if they don't. These are very expensive items. I would imagine you would keep track of everything > > Dr. Vinci Chow Recapping the ridiculousness of the situation, Chow stated, “All these high-end GPUs are still manufactured in China, but they're not allowed to be sold in China. And somehow the US government thinks that's going to work and somehow the Chinese government also allows that to happen. I have no idea how actually that whole thing works.” We asked both professors if NVIDIA knows whether all of this is happening and Chow stated, “I would be surprised if they don’t, right? I would be surprised. I would be really surprised if they don't. These are very expensive items. I would imagine you would keep track of everything.” * * * * ##### **Hunting for Banned GPUs (Golden Computer Center)** After this discussion, Zihao parted ways and Vinci brought us to one of his favorite tech spots in Hong Kong: The Golden Computer Center and the outdoor Apliu Street tech flea market. We asked him if we could go find some supposedly banned GPUs, like the RTX 5090, just available out in public. The hope was to find a shopkeeper with some ground-level or consumer knowledge. When we asked Chow if the export controls are having their desired effect for the US, he responded, “Well, yes, in terms of stopping China from building a comparable GPU cluster to the US.” While we were visiting Golden Computer Center, Chow pointed out some purchasable “parallel import” GPUs, which means they were smuggled into the country. We saw banned cards for sale here, including the 5090 Founders Edition. Despite the sometimes inflated prices, we were surprised at how easy these GPUs were to purchase. It seems like one of the main impacts of the GPU bans are the prices of the cards. Chow theorizes that the cards are moved into the country one by one, often by traveling students. We asked a shopkeeper there, “What’s the most popular card?” He said it was the 5090 and that they come from Australia and Taiwan, neither of which has an export ban on these GPUs to China. He also said they can “buy a lot” of 5090s from mainland China. Hong Kong was impressive and filled with character and culture -- and GPUs, apparently, because one of our next sources sent a message the night we were planning to hitch a ferry to Shekou Port in Shenzhen. He told us that we could buy a GPU as soon as tomorrow if we wired him money immediately. * * That seems like the responsible thing to do, and because it’s fully above board as a buyer, and because he texted us a photo of an RTX Pro 6000, which was very intriguing since it’s Blackwell, we decided the best way to learn more about buying export-controlled GPUs in China would be to just do it. So far, all the contacts were aware that some form of factory repurposing, theft from the line, QC rejects, and actual by-hand smuggling are involved -- but none knew for sure how the GPUs move. To get closer to the sources, we decided to wire the funds and get an address. We ended up sending out $3,289 via wire on a tight deadline, with our boat bringing us to the seller the next morning. ### **Turn a Blind Eye: How GPU Smuggling Works** #### **HUIZHOU: Mr. 5** We then packed up and boarded a ferry to Shekou Port, but we decided to take a quick detour: Rather than go straight to Shenzhen to meet our GPU plug, we first went to meet up with someone else -- a source who told us he has more information on how GPUs get into China. We traveled to HuiZhou to meet a guy known as “Mr. 5.” HuiZhou is a city with some serious grit. * * * The city has grown to take some of Shenzhen’s factory industry as it’s been pushed out over the years with Shenzhen’s expansion and as it’s turned into more of a metropolis. We spent a good amount of time in HuiZhou over the last decade, mostly visiting case, painting, tempered glass, and tooling factories. "we asked him how 5090s get into China. He responded, 'It’s like this. China already produces the heatsinks and components. China makes a lot of 5090s. That’s the first way in.'" Mr. 5 is a cooling hardware reviewer and has specialist knowledge in factories that make video card cooling solutions, including NVIDIA’s. This experience allows him more access to information about the peculiar relationship between factories making cards and the companies that technically can’t sell them to the country where they’re made. We instantly related to Mr. 5 for his own editorial disputes with NVIDIA’s review sampling process, where we have felt the company seeks to control review direction. > They can buy it and use a ‘human flesh backpack.’ They carry that back and it’s not illegal in China > > - Mr. 5 His username is “51972” on Bilibili. ##### **“Human Flesh Backpack”** Speaking to him in Chinese, we asked him how 5090s get into China. He responded: “It’s like this: China already produces the heatsinks and components. China makes a lot of 5090s. That’s the first way in. The second is [...] The US has a ban on sales to China, but there’s no ban in China. Because Shenzen and Hong Kong are close, there’s only one customs check. Many people can get it from Hong Kong or other countries. For example, America, Japan, Singapore, etc. They can buy it and use a ‘human flesh backpack.’ They carry that back and it’s not illegal in China. They can purchase it that way. There’s one other method: for some people, it’s just for money and they choose to smuggle. So they use a ‘special channel’ (smuggling) to get it back.” > A few months ago, if you carried an RTX 5090 in from outside, you could earn about 2,000 to 5,000 RMB. You could make so much money. > > - Mr. 5 We asked Mr. 5 if he thought most of the banned GPUs come in one at a time, and he replied: “There are many ways. In China, we have a special type of ‘job,’ scalper. He can go back-and-forth many times in one day to bring them back, and every time he does, he brings some back. He can also organize people or a group of people to go to Hong Kong together to buy. Then he gets the difference in price. Each card gets from 500 to a few thousand RMB. That’s the method he uses. Strictly speaking, this kind of action is illegal, but it’s a gray area. They can organize people from Shenzhen, Hong Kong, or other countries. Of course, there are many other ways, like international students bringing them back. In China, the 5090 is not banned.” He added, “A few months ago, if you carried an RTX 5090 in from outside, you could earn about 2,000 to 5,000 RMB (about $280 to $700 USD). You could make so much money. The craziest is when the 5090 released during Chinese New Year, some people who brought one back from overseas made 10,000 RMB (about $1,400 USD). So a lot of people thought, ‘If I go abroad, I’ll buy an extra one.’” > Of course [NVIDIA] know. For this matter, of course they know, but, how, ‘how do I say this?’ ‘Open one eye, close one eye.' > > - Mr. 5 Mr. 5 stated, “China has an old saying, ‘All the hustle and bustle in the world is only for money and interest.’ I also have to add that many of the RTX 5090s are ‘made in China.’ The video card was born here locally, so a few brands choose to sell domestically in order to digest inventory. Because it takes a few months to ship by ocean freight--we already know it takes 1-2 months to get from China to the USA by boat, which wastes time. But if it’s sold domestically in China, it can be turned-around quickly. Funds return faster. This is also lower pressure and reduced inventory.” ##### **NVIDIA’s Awareness of Smuggling** When we asked Mr. 5 if he thinks NVIDIA knows about what’s going on with all of the banned GPUs being sold in China, Mr. 5 replied: “Of course they know. For this matter, of course they know, but, how--how do I say this?--‘Open one eye, close one eye.’” This translates to “turn a blind eye.” When we asked Mr. 5 if NVIDIA would want to stop it, he simply replied, “No” since the Chinese market is so big. ### **SHENZHEN: Suppliers, Middlemen, & Fixers of Banned GPUs** * * Our next stop was in Shenzhen Bao’an to meet with a trading company that sits between Hong Kong and Shenzhen warehouses. They’re the trading company’s trading company, and we planned to buy an RTX 5090 from them. Shenzhen was a fishing village just 40 years ago. Now, it’s one of the most technologically advanced cities on earth. Shenzhen has some extreme Cyberpunk vibes with its mix of technology and surveillance. * * It has packed tech markets in Huaqiangbei, including some of the weirdest computer parts we’ve ever seen. There are entire buildings dedicated only to small phone repair shops with their own kiosks, another building dedicated to a mix of computer hardware, gaming, and miscellaneous components, more still for just integrated circuits, and all of these places have people who know people. That’s their job. But with Shenzhen’s technological rise comes with it an uncomfortable omnipresence of CCTV and facial recognition, which feels more fitting today than ever before now that we know worldwide government facial recognition is a big user of AI GPUs. ##### **Buying an “Illegal” GPU** We started the day by meeting up with our translator for the next two days, Raymen, who’s helped us on factory tours for years now. Because Uber doesn’t work in China -- since it uses the Google Maps API and Google is blocked by the firewall -- Raymen also was our man with the locally compatible apps to get us places. This became especially important now that even cash is becoming less acceptable by some cafes and restaurants, as everyone has moved to paying with the WeChat app. With our ride booked, we set off to see if the wired funds turned into a GPU and tried meeting up with our GPU supplier. Upon meeting with them, we explained that we were media, and they allowed us to record them talking but didn’t allow us to show their faces or divulge their names or company name. * * They gave us the RTX 5090 we purchased, which wasn’t even much more expensive than what we would have paid for it in the US. We asked them what the most common GPU is that their customers buy and one of them, whom we’ll call “Silver,” stated the H800, A800, and 5090 D. These are are all banned GPUs (the 5090 D was not originally banned). We then asked them what sells better between the 5090-class cards and the H100 type GPUs, and they replied “5090 D.” We asked them if anyone buys high-end AMD GPUs and they responded, “Very few. AMD’s GPUs are rarely useful. We’ve had second-hand customers, but high-end is very rare.” We then asked to see if people were buying Intel GPUs, and they responded, “Intel is the least!” > '睁一只眼, 闭一只眼 '/ 'Turn a Blind Eye' > > - "Silver" We then asked them if they think NVIDIA knows about people buying and selling banned GPUs in China and they echoed what Mr. 5 said and replied, “They ‘open one eye, close one eye.’ They know you can buy it.” The most important thing we learned here was where to go next: Just like how the ants move GPUs piece by piece, we’ll have to collect our information piece-by-piece. The company said that some cards have become more difficult to get, but they can still get them. The fact that they texted us a photo of an RTX 6000 PRO Blackwell card -- the very same that we just bought for $8,500 in the US and that is hard to get at home -- shows that they’re resourceful. Their price is $8,600 US, which is actually cheaper when factoring-in US taxes and shipping. So the ban isn’t stopping them, and the 6000 PRO Blackwell GPU is a serious AI card (that we’ve benchmarked) with 96GB of memory. It’s banned and in demand. This company isn’t used to single-GPU sales like ours: It mostly transacts with other trading companies and in high volume, interfacing with a Hong Kong distributor to bring the cards into Shenzhen, moving it across one more border, another ant in the chain. That meant they could connect us with their distributor, which would be familiar with where we could find smugglers, but this company itself neither knew many smugglers directly nor knew many end users. That’s OK, because each link in this chain will get us one ant closer to the information we need. We took note of their information on the distributor and smuggling side to use when we got back home, then they helped point us toward our next stop: Huaqiangbei, which is located in Shenzhen. We hopped in a car and drove 40 minutes to meet with a GPU trader we found in the city, over near Huaqiangbei. #### **SEG E-Market: The Biggest Tech Market in the World** Huaqiangbei is home to the world-famous SEG E-Market, or Saige, where we’ve found some of the strangest computer parts we’ve ever seen. Huaqiangbei has the highest concentration of technology and integrated circuits in the world, with a neverending maze of multiple disconnected, multi-story malls specializing in all electronics. If your life depended on getting a complete product made in a single city block, Shenzhen Huaqiangbei is your best bet. Fortunately, this time, we’ll be with this guy: This is Vincent. His profile picture on one messaging app is Van Gogh, so he has a sense of humor. He seems to have a natural ability to make people -- and cars -- get out of his way and he argued with security guards about us being able to film. Vincent generally seems to have an attitude of getting shit done. But before we met him, we had some concerns going into this one that it’d be fruitless or that no one would even be there to meet us. Luckily, he met up with us after we arrived. > We can only move [GPUs] slowly, but there are a lot of people in China. > > - Vincent We asked Vincent if it was difficult to get the GPUs. He replied, “The market has a lot of them” and pointed us towards the nearby SEG E-Market. Vincent runs a trading company and buys and sells the GPUs. We asked him how the high-end GPUs get into China, and he said, "They often come through Hong Kong” and added that “Taiwanese people also sell them here.” He confirmed that the GPUs are often brought over one-by-one. When we asked him if this one-by-one movement was enough, Vincent replied, “We can only move them slowly, but there are a lot of people in China.” > If it’s a common card like a 5090, we have them in China. Chinese factories made those [5090s] and sell them > > - Vincent ##### **Consequences of Smuggling** We then proceeded to ask if there were any consequences from bringing those cards in. He replied, “If it’s a banned video card, then China doesn’t have any first-party [cards]. You can only bring it in from outside. If it’s a common card like a 5090, we have them in China. Chinese factories made those [5090s] and sell them. That’s the way it is. If you’re talking about high-tech servers, you can only bring them in from outside. China doesn’t have these [high-tech parts].” When we asked Vincent if China cares, he responded, “It’s not the Chinese government’s business. It’s a US ban.” He revealed that “each person has their own method to get cards in. Normally, I just get it from Hong Kong because it’s close.” We asked Vincent what his normal order amount was and he replied, “Relatively low. Just 1 or 2 [per customer].”  We then asked what was the most in-demand GPU, and he responded, “The 4090 is relatively popular, but the 5090 is too expensive. No one wants it.” When we pointed out that the 4090 was banned, he was surprised and stated, “Really? 4090? It’s relatively common.” We found it interesting that some dealers of "banned" GPUs aren't even aware that they are banned, illustrating just how easy these parts are to get. Vincent stated, “We don’t know the reason for the ban, we can only buy and sell what’s on the market.” We then asked Vincent if we could follow him to SEG E-Market and walked there with him. Vincent led us over to various large wholesale and consumer retail markets. His office is within about a 10-minute walk from SEG E-Market, which is his daily haunt. There’s one peculiar detail in all of this: Vincent has customers both in and outside of China, but his foreign customer base is relatively high. He mostly deals in RTX 4090s, which seems to be a trend as you’ll see with our smuggler contact at the end. That means that Vincent is buying 4090s that either never left China -- but were supposed to -- or were likely illegally re-exported to China, and is then re-re-exporting them to his foreign customers. It’s like an infinite loop. He used to work as a shopkeeper in the mall itself and would interface with factories or anyone else who wandered in. > we went into a shop and asked if they had 5090s. The shopkeeper said that they do These days, he’s closed shop and is a buyer instead. He mostly takes orders online, doesn’t keep much inventory, and then just walks across the street to buy whatever was ordered that day. It’s basically a concierge GPU-picking service. While at SEG market, we went into a shop and asked if they had 5090s. The shopkeeper said that they do and they confirmed that they did have them and could sell them for 19,800 RMB, which was roughly $2,757 USD. That’s not bad and is pretty close to US pricing. Speaking to shopkeepers, and like with the other interviews, we learned that people rarely want AMD GPUs and basically no one wants Intel. Along the way, he introduced us to his cousin in the integrated circuit business. > Vincent walked us around and repeatedly expressed confusion at the ban list. He sees RTX 4090s everywhere and RTX 5090s are easy to get His cousin connected a missing link for us: People like him, selling FETs, inductors, capacitors, resistors, MLCCs, and so on for repair shops, can help to supply the parts needed for upfit or modifications. For GPUs, this can also include sourcing additional VRAM for modifications. > some sellers told us that they’re capable of sourcing devices like the A800 or even A100 Vincent walked us around and repeatedly expressed confusion at the ban list. He sees RTX 4090s everywhere and RTX 5090s are easy to get, just expensive. He was pretty sure that their abundance meant we were mistaken, but we checked and re-calculated, and they are export controlled devices. He told us about how 5090s are easy to get because the local factories supply them, matching Mr. 5’s comments earlier. Vincent pointed out everything from RTX 20-series GPUs, GTX 16 cards, 30-series, A1000 cards, Quadros, BTC mining rigs, and 5090s, and 4090s. Although some sellers told us that they’re capable of sourcing devices like the A800 or even A100, they did not have them on-site. They have them in a more secure spot. That’s because there’s a separate area for the warehousing. Fortunately, we learned where they are from our new friends in SEG. Like every other link in this chain, we kept getting closer to understanding the full story. Grab a GN15 Large Anti-Static Modmat to celebrate our 15th Anniversary and for a high-quality PC building work surface. The Modmat features useful PC building diagrams and is anti-static conductive. Purchases directly fund our work! (or consider a direct donation or a Patreon contribution!) ### **Warehouses** On a rainy morning, we headed over to the warehouses containing all of the GPUs to get a better idea for how many there might be. SEG’s guards have never been particularly friendly toward filming to begin with, but ultimately, they’re not police. The worst that happens is we get asked to leave. Besides, there weren’t any signs saying we couldn’t be there. > From the perspective of the United States, the warehouse we visited is filled with highly illegal items. They didn't have export licenses. > > From the perspective of people working here, however, they’re just doing their job. It's not shady. It's not some subterranean refuge of firearms. It's just a poorly lit warehouse like any number of other warehouses around the world, but the perspective of the American government would be different from the people working here, who just want to sell these things and go home. * * * * If GPU sellers don’t have certain GPUs, they purchase them from these warehouses. We saw stacks upon stacks of GPUs for sale. These feed into SEG and get distributed to Alibaba sellers like Vincent and get sent out from there. To be clear, not all of the GPUs we saw in the warehouse are banned, but we did see a couple 5090s. From the perspective of the United States, the warehouse we visited is filled with highly illegal items. They didn't have export licenses. From the perspective of people working here, however, they’re just doing their job. It's not shady. It's not some subterranean refuge of firearms. It's just a poorly lit warehouse like any number of other warehouses around the world, but the perspective of the American government would be different from the people working here, who just want to sell these things and go home. The warehouse security did eventually kick us out for filming, although they were oddly polite about it, which we appreciated. ### **Zhengzhou** Next up, we headed to Zhengzhou on a 5 hour and 50 minute high-speed rail ride to visit Brother Zhang’s repair shop, where they build custom 48GB GPUs. * * * * * * Brother Zhang is famous for repairs on Bilibili. He’s got video cards and spare parts all over his shop and he saves everything that he can salvage, then rips the good components from the boards to reuse them for other ones. Brother Zhang is an interesting guy in the chain because he’s not directly part of the GPU black market -- he just fixes video cards. He does that for consumers and for companies. Sometimes, that means people want him to modify video cards to, for example, double the VRAM. That’s where it gets interesting for our story. Speaking to him, we learned that his shop tests about 50 GPUs a day and, while we were there, we saw them testing 4090s and 5090s. Most of the GPUs are from NVIDIA, but there are some from AMD, with relatively few from Intel. Most of the cards that come into his shop come from all throughout China and most of his customers know him because of his videos. "He converted a 24GB RTX 4090 into a 48GB RTX 4090" The ultimate reason we’re at Zhang’s shop is to see how video cards are modified. Because if there’s an export ban on GPUs going to China, then it becomes critical for China to be more self-sufficient in keeping those GPUs that they do manage to get to stay in service. Most of the repairs are more typical things you'd expect of soldering. There's board heaters, soldering irons, surface mount components gathered from suppliers nearby, and a lot of test stations. But one thing that's unique to China and especially to this shop is the ability to take an existing model video card and completely modify the SKU into something that NVIDIA doesn't even make. NVIDIA is intentionally restrictive with how much RAM it puts on cards. Part of that is to upsell people to more expensive models. Actually, that's pretty much all of it. That's basically why they do it. But at shops like Zhang's, they double the VRAM on some of the cards that come through, making it into a fully custom SKU. We wanted to see if he had one lying around that he could show us, and he did one better than that: his shop made a modified 48GB RTX 4090 right in front of our eyes. * * * He converted a 24GB RTX 4090 into a 48GB RTX 4090. The unit that we looked at went through 6 repairs: It originally didn’t work at all, then it had a memory problem, then it had display issues. While the card works now and the GPU itself is functional, because its board has had so many issues, it would be better to harvest the expensive components (like the GPU) and put them on a new board to avoid other potential failures. When we asked Zhang what the success rate was, he told us it was about 99%. Impressive. Because NVIDIA doesn't make a 48GB 4090, the repair shop has to source its own PCB and cooling solution. Other shops commonly sell both of these on the open market in China. It needs something that has enough pads for all the memory modules they're adding to it, more memory, and a pre-populated VRM. They work with a third-party supplier that builds a PCB, uses an SMT line to place all the VRM components, and has the extra wiring, circuitry, and pads to support 48GB of memory. * * * To create a 48GB 4090, the first step is to disassemble the card. * * * From here, they heat the GPU to 260 degrees Celsius for about 5 minutes, which allows them to pull the GPU off the PCB. Zhang revealed to us that they get about 10-20 orders at a time and, when we asked him if NVIDIA had contacted him to tell him to stop, he responded, “I don’t think they will.” * * * * The next step is to work on the memory. To do this, they place a PCB on a board heater and use a hot air station to heat pinpointed memory modules and remove them without risking damage to other components on the board. From there, they’ll add them to a newer board. This, ultimately, allows them to double the card’s memory capacity. * * * Next, they took out a template for the solder balls and cleaned it with rubbing alcohol. From there, they poured solder balls into the template and positioned it on top of the memory modules. This allowed them to bake the new solder balls onto the memory modules, which they could then attach to the PCB. They then applied flux to wet the solder balls and used a solder wick to pull off the excess solder. > From start to finish, it took [the shop] about 2 hours to make a 48GB 4090 When we asked Zhang what 48GB RTX 4090 cards normally sell for, he told us over 20,000 RMB, which is roughly equivalent to $2,785 USD. That’s not a bad price. * * * * * From there, he put a lid on the template and poured in soldering balls, allowing them to adhere to the memory. They then heated the modules on a hot plate that ran at around 195 degrees Celsius. Because the template only allowed them to do 8 at a time, they had to do the solder ball process twice. They then applied heat directly to the surface where the solder balls were using a hot air station. * * * * The next step involved cleaning the GPU in order to mount it to a new PCB. From there, they added flux to the GPU. They then applied a custom template to the bottom of the GPU and poured solder balls on top of it and used the template to sift the solder balls. * * * The next thing the shop did was place the GPU on top of a jig to hold it in place on top of a hot plate to protect it. They then placed the memory modules onto the PCB and, again, used the hot air station. * * Finally, after about 2 hours, the shop added the GPU to the PCB and placed it under heat to bake the GPU back on. * * * * From there, the team cooled the GPU off with a fan and added thermal paste. A technician placed a heat sink on top to test the card to see if it worked, leading to an instant success. * * * * From there, they attached the GPU and PCB to the rest of the video card’s cooler and chassis. From start to finish, it took them about 2 hours to make a 48GB 4090. At the beginning of this process, the card was working, but it had been through so many repairs that it was very likely to break again at some point from something they couldn't really predict. From a waste standpoint, it's just better to figure out a way to save that card because the PCB is kind of the least valuable part and is also the least precious of the resources. Being able to save the silicon that's in the memory and in the actual GPU itself provides a lot of value, but it especially offers value in China, where they just don't have as much supply, even though the supply looked good from what we saw. Purely from the perspective of people in China, not only is this better than what NVIDIA shipped to begin with, and really not that far off in price when the card was at its most scalped, it's also just a good way to keep the silicon in circulation even as things like MOSFETs, capacitors, or PCBs die. The silicon tends to be pretty resilient and it doesn't really die that often on GPUs. What this tells us is that this is maybe an alternative method to getting more GPUs into supply. If they take the broken boards and then put functional silicon back into circulation, it's a certain level of ingenuity, and there aren't many places that are set up to handle it. Brother Zhang's shop is impressive. ### Video Card Factory Next, we headed back to Shenzhenbei in a high-speed rail trip that added roughly 1,000 miles to our journey, planning to meet up with Yeston Creative Director of Product Zhou. Yeston is a video card manufacturer. * * * * * * * Zhou and Yeston were not officially contributing to the black market side of our research, but they did show us around their factory and there were a few interesting things we learned that relate to our story. The big takeaway is that factories receive their GPU supply from their partners. That would be AMD, Intel, or NVIDIA, and the memory is often packaged with it. But for the most part, everything else that they use, unless they're buying reference PCBs, all come from whatever sources they want to find for their supply. > for at least the gaming class cards, which the 5090 more or less is one of, there's enough volume and NVIDIA’s hands off enough that it would be pretty easy to make these disappear We were also informed that there's very little oversight in terms of the management of the rejects at these places. They can file for refunds if, for example, a chip is bad. But if they screw up a board in their own process, then it's going to be on them to deal with the defect. Factory work for GPUs is highly automated. For at least the gaming class cards, which the 5090 (more or less) is one, there's enough volume and NVIDIA is hands-off enough that it would be pretty easy to make these disappear. The biggest reason for this is that there are partners for the 5090s, unlike cards that only NVIDIA makes, such as the RTX Pro 6000. Since there are partners, NVIDIA is only tracking when they sell those partners the GPUs. NVIDIA doesn't necessarily keep tabs on where those get distributed once they're done being manufactured. For this reason, if a factory wants to just sell the GPU domestically in China and they think they can get away with it without being added to the entity list, which would very likely kill their business with NVIDIA, then they might sell them domestically. That might be how some of these trading companies get a hold of them. It's an easier way to make them disappear than through channels that are more publicly visible from NVIDIA and the government. Yeston doesn't make 5090s, but the processes they taught us about illustrates how that could happen. As for the server grade solutions, there are manufacturers for those that are pretty tightly controlled. Basically, the silicon is made in Taiwan, shipped to wherever it's going to be assembled, and accompanied by the other components sourced in China. As a result, it's easiest to keep the whole supply chain there, including the assembly. > after seeing how a highly skilled group like Brother Zhang's repair shop can just repurpose all those components that are still good, you could see how there would be value in figuring out a way to make that board disappear, even if it's supposed to be in the trash Factories use surface mount technology (SMT) lines with conveyed inflow of PCBs and reels upon reels of components that are placed smallest to largest, with the most valuable typically at the very end. That would often be the GPU and VRAM. Heatsinks are installed often through a manual process at the end. There is often manual assembly even for NVIDIA's high-end products. We have seen NVIDIA's server video cards (and servers) being made in different factories. If the factory finds a defect, however unlikely, it could 'disappear' into scrap to be resold to the domestic market later. Dr. Vinci Chow had such a board where he had a broken link, and so there's a good chance that card stayed in China and never left. If it was reported as a defect to NVIDIA, then it may have effectively just been written off and considered trash. After seeing how a highly skilled group like Brother Zhang's repair shop can just repurpose all those components that are still good, you could see how there would be value in figuring out a way to make defects -- or "defects" -- disappear, even if it's supposed to be in the trash. This comes back to NVIDIA turning a blind eye, because they'd eventually notice it from the serial number popping up somewhere. NVIDIA might choose not to notice it, though. Besides, once someone has a card, especially in China, there's not a ton NVIDIA can do, anyway. ### **Taiwan** We then headed to Taipei, Taiwan. Most of our stay in Taiwan was for another story; however, we did accidentally unturn one stone with something interesting under it. A B2B company in Taiwan noted to us that it is commissioned by Chinese companies to import servers for pre-testing and pre-assembly and setup steps. When we asked what GPUs the servers ran, we realized that all of the hardware is export-controlled. Once the B2B agency completes its testing, it ships the system out to the original buyer. Basically, they act as an intermediary to buy the machine, then mark it up and reship it to the original purchaser. We also spoke with a company that conducts business in Singapore. The company informed us that they are also aware of similar passthrough so-called “testing” services. ### **Smuggler** Our last link in this chain is actually the first: The smugglers themselves, and we’re back in the US for this. We didn’t think we’d be able to find one doing the dirty work, but in the final hours producing this story, one of our viewers provided a lead. This led to about a day of production delay as we were wrapping this project up, but as far as we’re aware, this is the first content piece that actually features someone doing the highest risk dirty work. Everyone else we spoke to is in China. They’re safe from US retaliation. But large-scale smugglers get arrested, fined millions of dollars, and can spend years in prison if they were serious volume movers. A viewer contacted us to say that he’d connected with a traveling GPU buyer. The story went that this guy drives around buying specifically RTX 4090s. He doesn’t care about anything else, including 5090s, because those don’t sell as well to China. 4090s are ideal, he told us, because they can be modified into cost-effective 48GB models, like at Brother Zhang’s shop. The smuggler was extremely open with us in text messages written only in Chinese, but wasn’t open to us flying out to him. He did allow us to share footage of his car, however. > The Plug offers $2,000 flat per RTX 4090, which isn’t a bad price. He told us that he then finds ways to get them into China The Plug, as we call him, has an ATX test bench, motherboard, simple downdraft cooler, power supply, obviously a CPU, RAM, and an SSD, and gigantic battery in the trunk of his Prius. He also has at least one spare license plate in his trunk in addition to the one on the back of his car. We’re not sure why, but it’s better not to ask. People like The Plug post on Facebook Marketplace and other common online reseller forums looking to buy GPUs, just like anyone else in your city would. The Plug offers $2,000 flat per RTX 4090, which isn’t a bad price. He told us that he then finds ways to get them into China. He recently was scammed out of $5,500 of payments owed by a Hong Kong buyer that resells in Hong Kong, so lately, he’s been considering hand-carrying the devices in luggage instead. His margin is slim. He makes just under $300 US per GPU, not counting gas, potential hotels as he travels, and time. Our understanding is that some in his shoes will strip the GPU cooler off and ship just the PCB back to China, increasing their margin. Coolers are available in abundance where they’re made, so it can be cheaper to do this than to pay for the weight and size. He sometimes buys entire computers just to take the 4090s out, as that’s what the bounty is on. He then sells the remaining system back to anyone who’ll buy it. He told us that 5090 prices in China are falling so fast due to oversupply, ironically, that he’d lose money reselling it to China and would do better flipping it to Americans. The Plug isn’t rich. He seems to be doing OK, but it’s not the type of wealth you might expect for such risk. We can’t overstate how important it was to get this piece into the story, as this allowed us to fully complete the chain. We want to provide a huge thanks to our anonymous viewer for their help. Many of the sources we met in Asia during this trip told us they simply didn’t know how the GPUs actually get in, and now we know. People doing this on a small scale like The Plug are unlikely to be caught, but operations transacting millions of dollars worth of GPUs would have a harder time getting them out undetected. That brings us to the end of our travels, but not the end of this piece. We still have two more entities to look at, and that’s NVIDIA and the US Government. ### **Details: Legality** The legality is simple -- this is our understanding of it. There is no restriction on purchasers, only on sellers the US has control over. Even Americans buying GPUs in China are not violating any laws, as the purchase of a GPU in China is not restricted. We’ll start with selling: The simplest answer to “who the sale of GPUs is restricted for” is anybody who doesn’t have a specific export license and who would be governed by US law, and NVIDIA can’t just bypass it by shipping from a different country. It has to do with the sale, not with the shipment location. That includes entities buying from US companies, like European companies, who would be under guidance both from NVIDIA and, if they do business in the US, its government. It’s also illegal for anybody in the United States, citizen or not, to sell these GPUs to China, Hong Kong, Macau, or companies in those locations if the seller does not have a re-export license. It is not illegal for a person of any nationality in America to sell a GPU to any non-restricted entity. As for buying, as Dr. Vinci Chow stated, “There’s absolutely nothing illegal for us over here to buy these GPUs.” But it’s also not illegal to sell them in China for a Chinese company. The Chinese government doesn’t generally enforce American export laws. Other nations cooperating with the United States might, such as the recent arrests in Singapore, but once the GPU is in China, the people in possession of it likely don’t care -- buyer or seller or someone who does both, like GPU dealer Vincent. The only control over Chinese companies that the US has is the Entity List, which would hurt their business prospects with American companies -- but only if they care about that. Chinese GPU middleman 思騰合力科技有限公司 (Sitonholy Technology Company) in Tianjin landed on the Entity List for transacting banned GPUs and being found out. Sitonholy purchases GPUs from anyone who can get them into China, and then they bid on domestic projects like data center build-outs. The Entity List was used to restrict, for example, DeepCool previously, resulting in their American partners ceasing business with them for fear of frozen assets, audits, or collateral bans. This hurts companies like DeepCool that want to operate in the US and even shut down their California office, but for Sitonholy, they mostly want to do business with other Chinese companies, and so it’d have limited impact unless they wanted to expand to do business with Americans. Beyond the entity list, the US really has no control over what happens inside Chinese borders. That means the only point at which a GPU could feasibly be intercepted and a person arrested would be operating in the US or in transit to intermediary countries, such as Singapore, which may have their own export laws. The people buying and selling them within China are not breaking any of their own government’s laws, though. > We think the story of AI GPUs has become a story of corruption between governments, and the wealthiest company on Earth. ### **NVIDIA is Playing All Sides** And so we come back to NVIDIA. At every turn, it really looks like NVIDIA is playing all sides. If there’s enough money to be made, anybody is NVIDIA’s friend. We think the story of AI GPUs has become a story of corruption between governments and the wealthiest company on Earth. For example, on April 30th, Amazon-backed AI startup Anthropic called on the US government to increase export control restrictions to China. As part of a blog post, Anthropic said the government needs to improve its export enforcement to reduce smuggling. The company cited examples of chips being smuggled with “prosthetic baby bumps” and “live lobsters.” This upset NVIDIA, obviously, because NVIDIA doesn’t like restrictions on making money. So, NVIDIA shot back. Seemingly taking a page out of Trump’s playbook, NVIDIA essentially called this fake news and stated, “American firms should focus on innovation and rise to the challenge, rather than tell tall tales that large, heavy, and sensitive electronics are somehow smuggled in ‘baby bumps’ or ‘alongside live lobsters.’” * * * > We unearthed this official Hong Kong customs website detailing a white van busted driving on the Zhuhai-Macau bridge, filled with 280 kg of undeclared live lobsters and, yes, 70 smuggled GPUs #### **Gaslighting by NVIDIA** Anthropic isn’t telling tall tales--it’s right. A 2022 video previously showed a security check at Zhuhai port in Guangdong, not far from Shenzhen, wherein a woman with a prosthetic baby bump was shown to have been carrying CPUs and iPhones instead. The report made it to customs.gov.cn, stating that she arrived from Macau, a common go-between (similar to Hong Kong). This story got international attention in technical media and we reject the possibility that NVIDIA wasn’t aware of it. We found the official Chinese Government posting about smuggling from Macau, so there's public record. Imports to most of China are taxed, and so tax evasion coupled with smuggling will increase margin on the electronics rather than sharing it with one of the two governments. If you’re already breaking a US law, it seems some just go for a hat-trick and increase the profits. * * * As for the lobsters, that’s real, too: We unearthed this official Hong Kong customs website detailing a white van busted driving on the Zhuhai-Macau bridge, filled with 280 kg of undeclared live lobsters and, yes, 70 smuggled GPUs, complete with photo evidence. > NVIDIA’s response to call these 'tall tales' then is not only defensive, but serves to gaslight and grossly mislead, we think The Hong Kong customs itself calls this “zousi,” or smuggling, and notes a maximum sentence of 7 years. NVIDIA’s response to call these “tall tales” then is not only defensive, but serves to gaslight and grossly mislead, we think, and is tantamount to lying for sake of downplaying reality for its own benefit. But then this is a common NVIDIA tactic, including its dishonest approach to reviews that we’ve already detailed and its deceptive and we think false advertising of the RTX 5070 as being equivalent to an RTX 4090, which is provably and comically false. Another instance of NVIDIA’s “fake news” defense was following a July 24 publication by the Financial Times, reporting that more than $1B worth of NVIDIA’s AI chips had been smuggled to China. In response, NVIDIA, whose blind eye is turning an awful lot lately, downplayed the issue, and stated, “Trying to cobble together datacenters from smuggled products is a losing proposition, both technically and economically. Datacenters require service and support, which we provide only to authorized NVIDIA products.” That sounds like something a company selling support would say. And the first part doesn’t really match the whole “the more you buy, the more you save” assertion. It’s only a winning proposition by all of their prior years of statements. If your only option is a useless insufficient data center or a cobbled-together sufficient data center, then a cobbled-together one is still a winning proposition by comparison. It’s weird for the company to pretend that this isn’t worth doing. It’s worth lots of money. #### **Tech Exec Sycophancy** That same day, whitehouse.gov posted an article titled, “Wide Acclaim for President Trump’s Visionary AI Action Plan.” Near the top of the post, it highlighted a sycophantic quote from Huang, which read, “America’s unique advantage that no country could possibly have is President Trump.” And we could call that sycophantic if it was about any president. On August 5, NVIDIA got another opportunity to talk. The US Department of Justice announced it had arrested two people in California for smuggling “tens of millions of dollars’ worth of sensitive microchips used in artificial intelligence (AI) applications” to China. The BBC reported that court documents say the shipments included the NVIDIA H100 and RTX 4090. Rather than admit smuggling exists, NVIDIA downplayed the situation and stated, “This case demonstrates that smuggling is a nonstarter.” Except that anyone who made tens of millions of dollars before getting caught had a pretty good start, and so did their customers. This is a bullshit statement from NVIDIA that seemingly aims to downplay and deflect to reduce lawmaker attention on its monopoly. NVIDIA also said, “We primarily sell our products to well-known partners, including OEMs, who help us ensure that all sales comply with U.S. export control rules.” NVIDIA noted that, “any diverted products would have no service, support, or updates.” Again, this is not fully true. Our own sources in this story noted that, although something like an HGX system would be hard to service, a standalone PCIe GPU could be parted-out and covered under a separate warranty, even in China. > a government committee called Deepseek a 'threat to national security' and said Deepseek had used NVIDIA’s technology ##### **Fears of Deepseek** Meanwhile, the US began scrutinizing NVIDIA’s technology for getting into China, whether or not the company itself was directly involved. In a bipartisan report called “Deepseek Unmasked,” a government committee called Deepseek a “threat to national security” and said Deepseek had used NVIDIA’s technology, “AI model appears to be powered by advanced chips provided by American semiconductor giant NVIDIA and reportedly utilizes tens of thousands of chips that are currently restricted from export to the PRC.” It continued, “NVIDIA designed and manufactured many of these chips to create the most sophisticated possible chip while skirting U.S. export controls. This has allowed these chips to be exported to China as the U.S. government develops stricter restrictions. Since March 2024, it is estimated that NVIDIA has produced over 1 million chips for the Chinese market.” ##### **Singapore Suspicions** The government also examined NVIDIA’s significant revenue growth in Singapore compared to China based on SEC filings, particularly in years featuring restrictions. The government questioned “whether PRC customers are arranging for the diversion of sensitive chips that are reportedly sold through Singapore,” since revenue from Singapore had grown from almost nothing since 2021. NVIDIA has defended its sales to Singapore by saying that “Customers use Singapore to centralize invoicing while our products are almost always shipped elsewhere.” According to NVIDIA, shipments destined to Singapore were only 2% of the company's total revenue in 2025. But we also know that Singapore has made numerous arrests relating to GPU smuggling, so there appears to be some reason for the concerns, whether or not NVIDIA itself wants to turn a blind eye to it. ##### **Ignorance is Bliss** Despite the allegations, NVIDIA downplayed any smuggling of AI chips. In a video uploaded 2 months ago, Jensen Huang spoke on smuggling, and stated, “Governments understand that diversion is not allowed. And there's no evidence of any AI chip diversion.” Except that there is evidence of it -- not only in this very story, but in readily available reports online for years now. Huang continued, “Our data center GPUs are massive. These are massive systems. The Grace Blackwell system is nearly two tons. And so you're not going to be putting that in your pocket or your backpack anytime soon. And so these systems are fairly easy to keep track of, but the important thing is that the countries and the companies that we sell to recognize that diversion is not allowed and everybody would like to continue to buy NVIDIA technology. And so, they monitor themselves very carefully and they're quite careful about that.” This one is interesting. Huang is right that it’s much harder to smuggle Grace-Blackwell or Hopper HGX-class complete systems. Dr. Vinci Chow’s statements align with this when he said, “It's very hard to get like a full HGX system.” But it still happens. At least one of NVIDIA’s GPU and server customers in another country told us that they facilitate intermediary transmission to China and in fact showed us the server racks on-site in their facilities. We weren’t allowed to film them, but we saw them. A separate representative told us that document forgery through third-party countries can also disguise such transshipments. One middleman told us that an NVIDIA distributor gets parts into China; a downchain factory told us that NVIDIA’s QC rejects sometimes end up repurposed and kept in China, salvaging the GPU and VRAM and scrapping the rest; Dr. Vinci Chow told us that one of his own devices had a defective link on it, contributing to this statement. And when he asked him whether he thinks NVIDIA knows all of this is happening, he replied, “I would be surprised if they don’t, right. I would be surprised. I would be really surprised if they don’t. These are very expensive items. I would imagine you would keep track of everything, right? It’s hard to know what [a person] plans to do with all of these defective parts, but I’ll be very surprised that no one has ever thought of the possibility that, if it’s something so valuable, someone would come up with a use [for] even a defective one.” ##### **NVIDIA’s Hypocrisy** So in one set of statements, NVIDIA said that smuggling doesn’t really happen because the export controls work and keep partners in-line. But in another statement, Huang called the US export controls a “failure,” talking out of both sides of his mouth, we think. He spoke of competing Chinese GPU brands posing a threat to NVIDIA, “The local companies are very, very talented and very determined, and the export control gave them the spirit, the energy and the government support to accelerate their development. I think, all in all, the export control was a failure.” But the stakes, and dollar signs, for NVIDIA had increased. Jensen said NVIDIA’s market share in China had dropped from 95% to 50%, and, in NVIDIA’s May quarterly earnings before the H20 exemption and revenue share, Jensen Huang said the company’s data center business in China was done, “However, the $50 billion China market is effectively closed to U.S. industry. The H20 export ban ended our Hopper data center business in China.” Or, as Jensen said, the China market is worth one Boeing. Boeing is probably not the best example... ### **Conclusion** Grab a GN15 Large Anti-Static Modmat to celebrate our 15th Anniversary and for a high-quality PC building work surface. The Modmat features useful PC building diagrams and is anti-static conductive. Purchases directly fund our work! (or consider a direct donation or a Patreon contribution!) > The October 7, 2022 Biden Administration export controls had a stated goal to 'protect US national security and foreign policy interests' by implementing new export controls Let’s look back at the timeline once more between NVIDIA and the US government, highlighting NVIDIA’s relentless appetite for global dominance. The October 7, 2022 Biden Administration export controls had a stated goal to “protect US national security and foreign policy interests” by implementing new export controls restricting China’s ability to build high-end semiconductors, including for the development of supercomputers. There was an included goal of staving off the potential for China to develop “nuclear weapons and other military technologies.” Blocking the H100 and A100 led to NVIDIA creating an export-compliant A800 at about 70% of the speed of an A100 for the Chinese market. A year later on October 17, 2023, the US Department of Commerce updated its export compliance and restricted NVIDIA’s A800 chip as well along with the newer China-targeted H800. Weeks later on December 6, 2023, NVIDIA told reporters in Singapore that it would be working on another new chip that would comply with the US’ new restrictions. #### **The AI Diffusion Rule** On January 13, 2025, which was after Donald Trump won the 2024 presidential election, but a week before he took office, the outgoing Biden administration tightened export controls by introducing national chip caps for many countries, except for 18 allies. That’s the AI Diffusion Rule we spoke of earlier and would have gone into effect in May. NVIDIA, who had remained relatively quiet about the regulations up until this point, criticized the restriction and made an attempt to appeal to the president-elect, and reportedly stated, “It makes no sense for the Biden White House to control everyday datacenter computers and technology that is already in gaming PCs worldwide, disguised as an anti-China move. The extreme ‘country cap’ policy will affect mainstream computers in countries around the world, doing nothing to promote national security but rather pushing the world to alternative technologies. AI is mainstream computing – ubiquitous and essential as electricity. This last-minute Biden Administration policy would be a legacy that will be criticized by U.S. industry and the global community.” It seems like NVIDIA tried to set up an appeal to the president-elect, stating, “We would encourage President Biden to not preempt incoming President Trump by enacting a policy that will only harm the U.S. economy, set America back, and play into the hands of U.S. adversaries.” We already went over the million dollar dinner and ensuing ban-then-unban of the H20 chip. ##### **NVIDIA Persuades Trump Administration** > roughly 1 month after Jensen reportedly spent $1 million to eat dinner at Trump’s Mar-A-Lago estate, the US Department of Commerce confirmed that it will not implement the AI Diffusion Rule that NVIDIA campaigned against Later that month on April 30, Huang said this of Trump, stating, “Without the president's leadership, his policies, his support, and very importantly, his strong encouragement[…] frankly, manufacturing in the United States wouldn't have accelerated to this pace.” On May 7, roughly 1 month after Jensen reportedly spent $1 million to eat dinner at Trump’s Mar-A-Lago estate, the US Department of Commerce confirmed that it will not implement the AI Diffusion Rule that NVIDIA campaigned against and that was created under the Biden administration. The rule was supposed to go into effect a week later on May 15. This does not unban GPUs like the H100, 5090, B100, and so on. Following the termination of what was supposed to be a rule to address national security implications, the Department of Commerce, now under President Trump, stated, “The Biden AI rule is overly complex, overly bureaucratic, and would stymie American innovation. We will be replacing it with a much simpler rule that unleashes American innovation and ensures American AI dominance.” NVIDIA predictably celebrated the statement. The company, which has been begging to sell to China while also praising Taiwan’s importance, now took an America-first posture, collecting countries like Pokemon, writing, “We welcome the Administration’s leadership and new direction on AI policy. With the AI Diffusion Rule revoked, America will have a once-in-a-generation opportunity to lead the next industrial revolution and create high-paying U.S. jobs, build new U.S.-supplied infrastructure, and alleviate the trade deficit.” Job creation promises coming from this company, in particular, are interesting, but they are playing all sides consistently. Later that month on May 28, Huang spoke with _Mad Money_ host Jim Cramer. He stated, “When [Trump] rescinded the AI Diffusion Rule, it was a visionary move. It was a bold move, and he recognizes that there’s an AI race and we’re not alone. And he wants America to win.” #### **A Manipulative NVIDIA** June 23, via Reuters, an official of the US State Department, which didn’t reply to GamersNexus except the Department did send us 3 out-of-office auto responders, warned of DeepSeek military and intelligence operations and warned of the use of “shell companies” in Southeast Asia to circumvent export restrictions. The report mentioned that DeepSeek had “large volumes” of high-end H100 chips, which are banned in China. NVIDIA didn’t like that, responded to Reuters, and stated, “We do not support parties that have violated U.S. export controls or are on the U.S. entity lists,” adding, “With the current export controls, we are effectively out of the China data center market, which is now served only by competitors such as Huawei.” Turning a blind eye to the situation. NVIDIA added, “Our review indicates that DeepSeek used lawfully acquired H800 products, not H100.” 3 days later on June 26, The Information reported that DeepSeek’s next AI model has been delayed due to a shortage of NVIDIA AI GPUs in China. This directly contradicts Huang’s comments that export controls do not work. On July 4, Bloomberg reported that the Department of Commerce, which also did not reply to GamersNexus’ emails, was preparing new export controls on Malaysia and Thailand to reduce chip smuggling; interestingly, Singapore, which now comprises a significant portion of NVIDIA’s revenue, was not on that list despite being a known smuggling passthrough. > [Jensen Huang] really pulled off something few tech CEOs have managed. He played both Washington and Beijing and he won > > - CNBC Business News anchor Deidra Bosa On July 10, Bloomberg reported that Huang and Trump were scheduled to meet again ahead of the CEO’s planned trip to China. Days later on July 14, NVIDIA confirmed that it will resume sales of H20 chips to China with Huang stating, “The U.S. government has assured NVIDIA that licenses will be granted, and NVIDIA hopes to start deliveries soon.” On July 15, CNBC Business News anchor Deidra Bosa gave her synopsis on the situation and said, “[Jensen Huang] really pulled off something few tech CEOs have managed. He played both Washington and Beijing and he won.” She added, “Jensen has stayed disciplined and diplomatic with a clear message, and that is: NVIDIA’s dominance serves America’s interest.” > We write to express our deep concern over the recent decision to resume exports of NVIDIA’s H20 chips to China. As policymakers and professionals with a background in national security policy, we believe this move represents a strategic misstep that endangers the United States’ economic and military edge in artificial intelligence. > > - national security experts to US Secretary of Commerce Howard Lutnick On July 28, The Financial Times reported that the US Commerce Department was not going to make “tough moves” to tighten export controls to China. This is in spite of several congressional members warning the administration not to loosen the US’ export controls for AI GPUs. Several national security experts also voiced their concern by sending a letter to the US Commerce Department, which read, “We write to express our deep concern over the recent decision to resume exports of NVIDIA’s H20 chips to China. As policymakers and professionals with a background in national security policy, we believe this move represents a strategic misstep that endangers the United States’ economic and military edge in artificial intelligence.” > NVIDIA might have started off as a much more humble company, but it has become a savvy political player in a game that’s seemingly pay-to-win On August 11, via Bloomberg, Trump said he was open to allowing NVIDIA to sell modified versions of the company’s newest Blackwell chips to China. That brings us to today. > There’s no one better equipped to play that game than the most valuable company by market cap in the world #### **Most Valuable Company by Market Cap in the World** NVIDIA might have started off as a much more humble company, but it has become a savvy political player in a game that’s seemingly pay-to-win. That seems only fitting for a gaming company to be particularly good at pay-to-win games. There’s no one better equipped to play that game than the most valuable company by market cap in the world, now at $4 trillion, led by a man whose net worth is estimated at $148.1 billion. NVIDIA knows when to bite its tongue and how to effectively appeal to the ego of politicians of all parties and all countries. We think NVIDIA is playing all sides. We think it is greedy, manipulative, and carefully employs propaganda such as its use of the “fake news” playbook for news which is literally reality. But we don’t think NVIDIA has a particular set of beliefs beyond just making more money. We think NVIDIA will sell anyone out to make a buck. NVIDIA is in the big leagues now. Inside of one month, reportedly paying $1 million to a sitting President after which followed the unlock of $5.5B of lost H20 revenue, followed next by a 15% split of that unlock going to the US Government, is what raises these new questions of NVIDIA's integrity in our piece. As for the actual black market side of it and smuggling, it was an exciting story to cover and get to the bottom of. We've learned that common methods include factory so-called QC defects, hand-carried items by students, actual smugglers on the ground in the US, and suppliers through third-party countries, among others. We loved working on this story and meeting all of these unique people. Each person played a key role in helping us find the next person, allowing us to complete the first public, on-record, complete start-to-finish cataloguing of a smuggling pipeline for high-end silicon. We want to thank everyone who made this story possible, including our viewers who funded it. * * *

Best CPU Coolers We've Tested (2025): Thermals, Noise Levels, & Value | 26 Coolers Testedjimmy_thang December 2, 2025 ## Our CPU cooler comparison includes benchmarks of the best air and liquid coolers (AIOs) for thermal performance and noise levels, spanning best overall, best budget, and more categories The Highlights * Best Overall CPU Cooler: Sudokoo SK700 * Best Noise-Normalized Thermals (Liquid): Tryx Panorama 360 * Best Noise-Normalized Thermals (Air): Noctua NH-D15 G2 * Best Budget CPU Coolers: ID-Cooling FX 360 Pro and Assassin Spirit V2 * Best Out-of-the-Box Thermals: Tryx Panorama 360 and A720 #### Table of Contents * AutoTOC Grab a GN Soldering & Project Mat for a high-quality work surface with extreme heat resistance. These purchases directly fund our operation, including our build-out of the hemi-anechoic chamber for our acoustic testing! (or consider a direct donation or a Patreon contribution!) ### Intro We’re rounding up some of the best CPU coolers that we’ve tested recently. Earlier this year, we completely overhauled our CPU cooler testing methodology to move to AM5 for the 9800X3D and 9950X3D. This round-up includes the best liquid and air coolers that we’ve tested this year and is a continuation of our Best Of series, which is one of our favorite video series each year since we can revisit the parts that are actually good. We just ran a story about the Best Cases of 2025 before this one, and both cases and coolers have had an overall positive year (in spite of the silicon industry’s best efforts, like with RAM, for instance). The round-up will include results from our AM5 platform, including 157W and 276W heat loads, both tested noise-normalized and at 100% fan speeds. _Editor's note: This was originally published on November 28, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication._ * * * #### Credits * * * ##### Test Lead, Host, Writing Steve Burke ##### Testing, Camera, Video Editing Mike Gaglione ##### Camera, Video Editing Vitalii Makhnovets ##### Writing, Web Editing Jimmy Thang * * * The categories this year include Best Overall, Best Value, Best Budget, Best Noise-Normalized (Air & Liquid), and Best Out-of-the-Box Thermals, which is just flat-out performance at 100%. ### **Overview** To set expectations on this content, our end of year Best Of round-ups are mostly meant for people who were out of the market for a while and are coming back in for a new build. We won’t go into as much detail as individual reviews, so we won’t be going through all the pressure maps and laser scans, and we’ll instead focus on the highlights. If you see something on this list that’s interesting, we’ll link full reviews below wherever we have them, or otherwise, check back for more cooler coverage soon as we get it all ramped and rolling out again. Product links are also below. For this round-up, we have mostly entirely new data that hasn’t been shown before, possibly with some content from older charts and reviews. The new data isn’t cross-comparable with older data, and in particular, our AM4 platform’s acoustic testing was done entirely differently from our new AM5 test platform. As usual, there are hundreds of coolers out there and we can only really speak to the ones we’ve tested this year and had hands-on time with. We have a good overall cross-section here, but if you don’t see exactly what you’re looking for, there are plenty of cooler reviews out there you could check for more detail. We’ll keep building on our list as well as we deploy this new methodology. Let’s get into the Best Coolers for 2025. Category| CPU Cooler| Review ---|---|--- Best Overall CPU Cooler 2025| Sudokoo SK700 on Amazon| Best Noise-Normalized Thermals (Liquid)| Tryx Panorama 360 on Amazon| Tryx '3D' Panorama Cooler Best Noise-Normalized Thermals (Air)| Noctua NH-D15 G2 on Amazon| Noctua NH-D15 G2 Review & Benchmarks, HBC & LBC Comparison, & Best CPU Coolers Best Budget| ID-Cooling FX 360 Pro on Amazon Assassin Spirit V2 on Amazon| Best Out-of-the-Box Thermals| Tryx Panorama 360 on Amazon A720 on Amazon| Tryx '3D' Panorama Cooler ID-Cooling A720 AD & TD, A410 TD, Cheap AIOs, & Scented Paste | Everyone is Targeting Thermalright Best Mechanical Design| Sudokoo SK 700V on Amazon| #### **Best Overall: Sudokoo SK700** **Sudokoo SK700:****Amazon** * The first award is for the Best Overall CPU Cooler for our 2025 testing. This award is given based on the total overall thermal performance, acoustic performance, build quality and mechanical design, and most importantly, value. Simply performing the best but being a $300 liquid cooler isn’t going to cut it -- we want it to be affordable for what you’re getting while also being compatible with most PC builds. This year, we’re giving Best Overall to the Sudokoo SK700 series. * We’re also giving a runner-up award to the ID-Cooling A720 (check out our coverage), which won last year. * * * The Best Overall category is always given to a cooler that won another category in the awards, and in this case, the SK 700V won our Best Mechanical Design award this year as well. As you’ll see in a few moments, Sudokoo is deeply familiar for its cooling roots, and its deep cooling capabilities aren’t entirely unique to this new brand. The SK 700V earns the Best Overall rank for a combination of its high build quality and its thermal performance. At 100% fan speeds for our 9800X3D workload, the cooler measured at 31.9 dBA while landing at 60.9 degrees Celsius for Tdie. That has it behind the A720 runner-up by just a couple degrees, but with what we think is significantly higher build quality at the same price -- and more importantly, better case compatibility due to the smaller stature and shorter 120mm fan. Even still, the A720 gets a runner-up award for Best Overall, maintaining some of its status from its victory last year in this category, simply for being the best not-$150-NH-D15-G2 air cooler on the chart. Maybe that’s a new award category we can introduce next year. The A720 is technically better than the SK700 series in thermals, but costs about the same right now and we do think its overall build quality is comparatively lower while also running louder in full fan speed tests (but with similar performance). * * * * That Sudokoo is achieving what it is while only running a single fan is impressive, and a lot of that comes down to the more densely-packed fin stack and its 7x 6mm heatpipes through the coldplate. The inclusion of the contact plate is also beneficial, but we’ll save that for the Best Mechanical Design section to get into. In our 9800X3D 157W heat load at 25 dBA noise-normalized, the SK700V landed at 64.1 degrees over ambient, so about 2 degrees warmer than the A720 Black. For this reason, the A720 Black still gets acknowledgement and a runner-up spot, but Best Overall is more than just cooling, and Sudokoo takes it with all that. * * We’re impressed with its ease-of-installation features, cleanliness of design, wide case and RAM compatibility, and thermal performance. DeepC--uh, Sudokoo has done well here. #### **Best Noise-Normalized Thermals (Liquid): Tryx Panorama 360 (Updated)** **Tryx Panorama 360 (Updated):****Original review****|****Amazon** The next award is simple and one of the most important. This is for Best Noise-Normalized Thermals, which is a purely objective measurement that ignores all other factors beyond the thermal performance given a set target of the cooler’s volume, so to speak. For this, we test the cooler in our hemi-anechoic chamber at a 1-meter distance, front-on, and adjust the cooler fan speeds until we hit a noise target of 25 dBA. We ignore all other factors, including price. Previously, this testing was done in a normal room instead of the chamber, so this is a big improvement for our accuracy this year. We’re giving this to two coolers: One for air and one for liquid. The first is for liquid. The Best Noise-Normalized Thermals (on liquid) award goes to the updated Tryx Panorama 360 (check out our coverage). In our updated noise-normalized benchmarking on the 9800X3D with a 157W heat load, the updated Tryx Panorama performed the best, but there’s a caveat here with dual-CCD CPUs. The result on this single-CCD 9800X3D was 53.6 degrees Celsius over ambient for average Tdie, slightly improving on the original Tryx Panorama’s result of 54 degrees and landing over 1 degree cooler than the Liquid Freezer series. In our new testing, because the acoustic chamber allows more accuracy than our older room measurements, the Liquid Freezer II (read our review) and Liquid Freezer III 360s (read our review) are often close by as a result of the impact to noise from the VRM fan. The VRM fan doesn’t affect CPU temperatures, but does affect the noise level, and so they sometimes exchange places depending on this. The pump in particular is noisy on the LF III, which we talked about in our original review. One of the worst 360s we tested was Cooler Master’s MasterLiquid 360L Core ARGB, which is a $60-$70 cooler. Their newer Atmos is better. The Thermalright Frozen Prism 360 (read our review) is cheaper (at $55) and outdoes air coolers, but not much else. With the 9950X3D at 276W and noise-normalized, the Panorama consistently performed slightly comparatively worse in the hierarchy than its result with the 9800X3D. This is related to the dual CCDs. Across 3 different test mounts, its performance was consistent -- but just slightly worse than with a single CCD. Breaking out both CCD results, the updated Panorama had some of the best CCD1 results, but performed notably worse on CCD2. Both results are fine, it’s just that the average works out worse. The Tdie result isn’t a clean average of just these two numbers (which would be 59.3 degrees), but is instead pulling from the HWINFO Tdie column. This comes down to pressure distribution and contact across the second CCD. * * * * Tryx’s Panorama cooler is best-known for its namesake 6.5-inch visible area, 60 Hz, 2240 x 1080 AMOLED screen, which wraps the cooler on one edge and can be used to play videos or display stats. We originally showed these at Computex when Tryx debuted it a few years back, but the company has been slowly iterating on the design since then. The display can also be used in a split-screen mode for two functionally independent panels. * * * This whole assembly makes the CPU block component tall and boxy, at 118 x 92 x 92 mm; although, the actual cooling block itself is a more “normal” height. Removing the display reveals a 60mm VRM fan on top of the block. Air primarily escapes one side of the fan through ventilated ports on the display block, although it’s able to exit out and down at least a little bit on two other sides. * * * Externally, the radiator is 30mm thick, but the finned area depth is reduced by about 9-12mm from that. That said, the extra spacing can help with reducing hub deadzones on the side with the fans. Tryx is buying from Asetek as its supplier, making it one of Asetek’s only recent relevant victories in the consumer space in years. The cooler is about $250 to $300, making it one of the more expensive ones on the market. You’re paying for that display. #### **Best Noise-Normalized Thermals (Air): Noctua NH-D15 G2** Grab a GN15 Large Anti-Static Modmat to celebrate our 15th Anniversary and for a high-quality PC building work surface. The Modmat features useful PC building diagrams and is anti-static conductive. Purchases directly fund our work! (or consider a direct donation or a Patreon contribution!) **Noctua NH-D15 G2:****Original review****|****Amazon** The next award is for Best Noise-Normalized Thermals with an air cooler. * * The other 9800X3D noise-normalized winner was the Noctua NH-D15 G2 (read our review), which managed to lead the A720 and Dark Rock Pro 5 by 1-2 degrees. The Royal Pretor (check out our coverage) didn’t do as well as we’d hoped. It was fine, but when noise normalized, it’s simply limited in performance due to its fans running on the louder side in combination with the lower speed impacting the 130mm fan more heavily. It does better than the Peerless Assassin in full speed testing, but not here. In the 9950X3D noise-normalized thermal testing, the NH-D15 G2 held a 69.4-degree average (over ambient), putting it at about 90 degrees once accounting for ambient. That’s borderline. It was the only air cooler in our test suite so far that was capable of safely completing this test without throttling or hitting a thermal trip. * * * * The NH-D15 G2 LBC is the low base convexity model, which performs best on AMD when compared to Noctua’s HBC and standard models. We demonstrated these differences in our original NH-D15 G2 review, including 3D laser scans of the NH-D15 G2 base plates with multipliers to illustrate the differences. We also took pressure maps of the D15 G2, illustrating the areas of highest contact and mounting pressure under the various coldplates. * * * * Noctua’s cooler is expensive. Repeatedly and on 3 platforms we’ve tested, it has had competitors within a few degrees of its performance that are around or over $100 cheaper, and that’s not counting liquid coolers as an option. It isn’t a good value. That much is clear. There are similarly performing coolers for far cheaper. Noctua gets credit though for its engineering, because by the performance and by the efforts invested into maximizing the coldplate and fans, the company has successfully iterated and advanced its performance. We said this category ignores all other factors (including price), and for that reason, Noctua’s NH-D15 G2 simply is the best for noise-normalized thermals on an air cooler. #### **Best Budget: ID-Cooling FX 360 Pro and Assassin Spirit V2** **FX 360 Pro:****Amazon** **Assassin Spirit V2:****Amazon** The next award is for the Best Budget CPU coolers. We’re giving these out in the liquid and air categories as well. For this award, a cooler needs to be as cheap as possible while still offering at least acceptable performance. This is for the price-conscious, and so we ignore almost all features aside from price and performance. * * For liquid, we’re giving it to the FX 360 Pro. For air, it’s the $15.59 Assassin Spirit V2. The $50 ID-Cooling FX 360 Pro takes it this year for liquid, dethroning the Thermalright Frozen Prism 360 of past years. These two companies have been locked in battle over cheapest coolers since DeepCool was evicted from the US market. Starting with the ID-Cooling FX 360 Pro. In our noise-normalized 9800X3D charts, it managed to at least run around the level of the other 360 liquid coolers we tested. * * * * * * The ID-Cooling FX 360 Pro is $52 currently. It’s cheap. Everything about it is cheap: The CPU block feels cheap and plasticky, the fans feel cheap, the stickers on the fans feel cheap, the unsleeved cables are cheap, and the simple mounting hardware and tubes are cheap. Even their website is cheap with its expired SSL certificate. That’s probably why it’s $52. But if you’re looking for cooling on a budget and don’t really care about the rest, then this can get the job done. * * * * The FX 360 Pro is all black and without LEDs, also helping contribute to price. As a positive, despite the cheap-feeling pump block, they’ve managed to get it to look OK in dim lighting in a case. The false brushed aluminum style plastic cap comes across OK visually. If you’re looking for a blackout build on the cheap, this might be a good fit. The cooler outperformed the Frozen Prism while managing to drop the price a few dollars lower, landing ID-Cooling in the winning spot for the Best Budget Cooler this year. Now the caveat: Liquid coolers have another angle to consider, which is longevity. We’ve covered a lot of liquid coolers over the years that gunked-up or had other issues. We can’t speak to the FX 360’s longevity since we haven’t had it that long. Currently, we have no immediate reason to suspect it’d be bad other than such low pricing always makes us a little cautious when we’re recommending products. * * * * As for the air category of Best Budget, it’s the Assassin Spirit V2 from Thermalright. We haven’t re-run these on our new charts, but we have tested them in the past. Like the FX 360 Pro, the Spirit V2 looks and feels cheap in just about every way, mostly because it is, but this is the best budget category. It’s $15.59 on Amazon right now. And you know when you have to charge 59 cents for something, you’re definitely pushing the limits of margin. The cooler can’t handle higher power CPUs, but for something like a 65W part, it’s capable. Even in our 123W 3800X testing we did years ago, it was capable of keeping the CPU at acceptable temperatures while on reduced fan speeds. * * * The Spirit V2 has 4x 6mm heatpipes with a direct contact base and a two-point mount. The fan maxes out at 1,500 RPM and is also a simple, cheaper solution without any frills or special engineering. The single-tower cooler gets its job done for about $16 though, so if you need something cheap for a lower power CPU or just to bridge the gap, or maybe get an older system up and running, this would work. #### **Best Out-of-the-Box Thermals: Tryx Panorama 360 and A720** **Tryx Panorama 360:****Original review****|**[**Amazon **](https://www.amazon.com/TRYX-SE-360-ARGB-Customizable/dp/B0F9WV43J1?tag=gamersnexus01-20)**A720:****Original review****|****Amazon** Our next award is for Best Out-of-the-Box Thermals. This is a simple award and is given for the cooler that has the best performance, regardless of acoustics. That means we just set the cooler fans to max speed and run as loud as they default to. These are often different winners since some coolers run crazy fast fans. For this, the best Out-of-the-Box Thermals at full fan speeds (with liquid) goes to the Tryx Panorama 360 again. Even when running full speed, it’s a chart-topper. By about 1 full degree Celsius, leading the Kraken and Corsair coolers in the 9800X3D chart. This was at 42.5 dBA, so it’s less acoustically efficient than the Kraken Elite when at max speed. The Corsair Titan is about the same noise level and within error of the Panorama. The MasterLiquid Atmos II is quieter at max speed, but also is warmer for CPU temperature. For air coolers on this chart, the winner is the A720, tied with the NH-D15 G2 and within error. The A720 has an advantage in its louder noise level than the NH-D15 G2, allowing it to brute force its way to an identical temperature result despite, of course, being louder. Our 9950X3D chart at full speed also goes to the Panorama, at 53.8 degrees over ambient. The Nautilus and Titan are close behind, as is the Liquid Freezer series. Cooling performance at full speed mostly comes down to how loud the companies want to run the fans, which has to do directly with speed; however, many of these entries are lower than winners of this category in years past. They’re still loud at full speed, but not as loud as something like the old EVGA CLC series and its odd fan choices of the time. #### **Best Mechanical Design: Sudokoo SK 700V** Visit our Patreon page to contribute a few dollars toward this website's operation (or consider a direct donation or buying something from our GN Store!) Additionally, when you purchase through links to retailers on our site, we may earn a small affiliate commission. **Sudokoo SK 700V:****Amazon** Our next award is for Best Mechanical Design. Years past have seen this awarded to the Noctua NH-D15 G2 for its impressive focus on mounting hardware and tuning to the physical aspects of the cooler, the DeepCool Assassin IV for its ease-of-use and installation features, and the NH-P1 (read our review) for its excellence in passive cooling. * * * This year, maybe fittingly with the Assassin’s prior victory, we’re giving the award to the Sudokoo SK 700V for its high build quality and fine-tuned mounting hardware kit. Interestingly, this is an AM5-only cooler. The Sudokoo cooler may look familiar, and that’s a special report story for another time, but the short of it is that if you’ve wanted a US entity list-compliant DeepCool cooler, there are options available. * * * * The Sudokoo SK700 is built of sturdy hardware that feels familiar to the old Assassin series we recommended for its build quality previously -- for maybe visually obvious reasons. The cooler uses an ILM replacement plate that secures directly to the CPU and helps clamp it down. This is like the Intel contact plates we’ve tested in the past, and it works well. When replacing the ILM, the plate is bolted down with 8 screws into the AMD backplate. We also noticed various offset mounting options for other or future Deep--uh, Sudokoo CPU coolers. * * * * RAM clearance is effectively infinite, and not only because people aren’t going to buy RAM at 500% higher prices than a month ago. The front fan adjusts vertically on a set of rails that are cleverly integrated into the sides of the tower itself. Plastic ridges nest with a metal frame mounted to the fan, allowing easy adjustability without the metal clips we often encounter, while also closing-in the cooler to force air through and out the backside. Depending on A/B testing, this could work better or worse for cooling, depending on the static pressure performance of the fan. * * The rails are cleverly done in a way that they could be swapped to any fan you wanted as long as it fits, resolving issues of CPU cooler fan rail systems in the past that required custom, one-off fans. This means that in the event of fan failure or just wanting to swap, you could migrate the rails to any fan that’d mount cleanly to the cooler. * * * Everything about the cooler has some heft to it, largely thanks to the use of 7x 6mm heatpipes embedded in a nickel-plated copper coldplate and a relatively high fin density of the heatsink. * * * * * The top of the cooler uses a pin-to-pad solution to drive power into the digital display plate, with two large access holes through the top center for accessing the mounting hardware with a long driver. Sudokoo even labels this to make it obvious, which is a nice touch for novice builders. All the plastic is removable as well with a few screws in the top of the frame. We like that the SK700 manages to pull together ease-of-installation features without making anything needlessly proprietary. * Our thermal testing in the noise-normalized 9800X3D benchmarks had it in the middle of the air coolers we’ve tested, behind the dual-fan Peerless Assassin while running only one fan, which is a great rank for it. The cooler is currently $60 to $70 (depending on SKU), making it competitive with last year’s Best Overall winner, the ID-Cooling A720. For all these reasons, the Sudokoo SK700 gets our Best Mechanical Design award. * * *

Best PC Cases of 2025: $75 to $200 Airflow, Thermals, Cable Management, & Budgetjimmy_thang November 29, 2025 ## We're looking at the best gaming PC cases for 2025 for custom PC builds. This round-up benchmarks the best airflow cases, tests for best thermals, acoustics, cable management, build quality, and more The Highlights * Best Overall Case: Fractal Meshify 3 * Best Mechanical Design: HAVN BF 360 * Best Sub-$100 Class Case: Lian Li 207 * Best Mid-Range Case: Corsair Frame 4000D RS * Most Innovative Case: HYTE X50 #### Table of Contents * AutoTOC Grab a GN Tear-Down Toolkit to support our AD-FREE reviews and IN-DEPTH testing while also getting a high-quality, **highly portable 10-piece toolkit** that was custom designed for use with video cards for repasting and water block installation. Includes a portable roll bag, hook hangers for pegboards, a storage compartment, and instructional GPU disassembly cards. ### Intro This round-up highlights the best cases we think are available in 2025, and it’s based on another year of extensive thermal benchmarks, acoustic tests, build quality analysis, and case reviews. The case collection has grown this year as we’ve now tested and reviewed hundreds of PC cases, but only a handful will get recognized today. The year had everything from bubbly, risk-taking cases to modernized retro-themed cases, with plenty of performance cases in between. _Editor's note: This was originally published on November 27, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication._ * * * #### Credits * * * ##### Host, Writing Steve Burke ##### Video Editing Vitalii Makhnovets Tim Phetdara ##### Writing, Web Editing Jimmy Thang * * * * * * * The good news is that there are a ton of really good cases right now, and this past year, in particular has been stronger for the cases we’ve reviewed than most others in our history. The bad news is that the entire rest of the industry is on fire (see here), but at least we have a sanctuary in cases. ### **Overview** : Best PC Cases for 2024 Category| Case| Review ---|---|--- Best Overall Case 2025| Fractal Meshify 3 on Amazon| A New Best: Fractal Meshify 3 Case Review, Thermal Benchmarks, & Noise Best Mechanical Design| HAVN BF 360 on Amazon| Our Most In-Depth Case Test Yet: HAVN BF 360 Flow Case Review, Fan Benchmarks, & Smoke Test Best Sub-$100 Case| Lian Li Lancool 207 on Amazon| Lian Li Lancool 207 Airflow Case Review | Cable Management, Build Quality, & Benchmarks Best Mid-Range Case| Corsair Frame 4000D RS on Amazon| Corsair Remembered How to Make a Case: Frame 4000D RS ARGB Review Best Noise-Normalized Case Thermals| Fractal Meshify 3 on [Amazon ](https://www.amazon.com/Fractal-Design-Meshify-Ambience-White/dp/B0CS3QXYZL?tag=gamersnexus01-20)Lian Li Lancool 217 on Amazon Antec Flux Pro on Amazon| A New Best: Fractal Meshify 3 Case Review, Thermal Benchmarks, & Noise Lian Li Lancool 217 Case Review: Thermal Benchmarks, Noise, & Cable Management Best Case of 2024 So Far: Antec Flux Pro Review & Benchmarks Best Out-of-the-Box Case Thermals| HAVN BF 360 Flow on Amazon Lian Li Lancool 216 on Amazon Lian Li Lancool 207 on Amazon Antec Flux Pro on Amazon| Our Most In-Depth Case Test Yet: HAVN BF 360 Flow Case Review, Fan Benchmarks, & Smoke Test Impressive: Lian Li Lancool 216 Case Review, Thermals, Cable Management, & Noise Lian Li Lancool 207 Airflow Case Review | Cable Management, Build Quality, & Benchmarks Best Case of 2024 So Far: Antec Flux Pro Review & Benchmarks Most Innovative Case| HYTE X50 on Amazon| The Weirdest Case So Far: HYTE X50 & X50 Air Case Review & Benchmarks Most Innovative Case Runner-Up| SilverStone FLP02| Unironically the Best Case: Retro Silverstone FLP02 with Turbo Button Genuinely, this year and last year have been some of the best years for cases we’ve seen in a really long time consecutively. This is awesome. There’s good innovation and interesting aesthetic designs that don’t sacrifice performance. For this story, we’re going to look at the best budget cases under $100, the best thermals, the best noise-normalized thermals, the best-out-of-the-box thermals, the most innovative, and the most mechanically complex cases over the past year. Specifically, we’re narrowing our scope to cases that came out towards the tail end of 2024 up until now. ### **Best Overall Case: Fractal Meshify 3  ** **Fractal Meshify 3:Original review | Amazon** Our first award goes to the Best Overall Case for 2025. Like last year, this was difficult to figure out for 2025 -- and that’s great news, because the case market is thriving with excellent advancements in design and performance. We talked through the many options: The Best Overall has to be universally good, including for thermal performance, acoustic performance in the chamber, value and price, build quality, aesthetics, ease-of-installation features, and cable management, and again, it all has to be priced reasonably. The Fractal Meshify 3 is our Best Overall Case for 2025, and in particular, the Meshify 3 Ambience Pro RGB (read our review) and the cheaper and less flashy Meshify 3 RGB. There are some better cases in individual areas, but we think this is a masterclass overall. We considered the HAVN BF 360 (read our review) for this category, but it just wasn’t technically the best thermal performer despite being good and is an expensive case. The Hyte X50 (read our review) was also in close consideration, as was the Lancool 217, but while they ticked many boxes, they didn’t tick all of them like the Meshify 3 does with its new price. * * * If not for the new price, the Meshify 3 would have lost it to one of these other cases. But now, the Meshify 3 Ambience Pro RGB is $130 to $144, a steep fall from the $220 launch price earlier this year. That instantly qualifies it for value consideration, especially with the $105-$110 price of the non-Ambience version (although we do think the lighting feature is key to the case). * * * The Meshify 3 Ambience Pro RGB brings back Fractal’s angular mesh front panel design and geometric pattern, using an ultra-fine mesh that breathes well while filtering dust as a byproduct. From some angles, it looks like a low-poly 3D render, and we mean that as a compliment. The way the light interacts with the case is well done, and Fractal’s use of the RGB lighting strips on its now $80-cheaper Ambience model shows care and attention put into something that we’ve memed on for a decade now, rather than just gluing an LED into a fan and calling it a day. * * * * * * * * * * The attention to detail extends beyond the LEDs, for those who care more about function: The integrated scoop in the power supply shroud actually helps in thermal performance, which we showed in our benchmarking. The rear foot of the case serves as a handle for the power supply dust filter, which is completely unnecessary but awesome attention to detail and good usability. This trend of attention-to-detail is nearly everywhere in the case, including deep cable management channels with durable plastic clips for cable management, included velcro straps at the rear of the case for securing external cables, and a motherboard tray cable cutout that runs the full length of the board to give freedom of routing without sacrificing tray strength. We also liked the panel quality, such as the top panel for its excellent removal and installation process, including a sturdy frame to reduce wobble. The front panel uses pogo pins to socket into the chassis for the LED strip, avoiding a wire. * * * * Clever functional features include flippable fan rail mounts that allow 120mm or 140mm fans to be mounted without either one being obstructed by the rails for the other. We see cases all the time that support multiple sizes, but block the flow of the larger fans with the rails that are present for the smaller ones. By using a mechanically clever and simple flippable rail, Fractal is able to leave both unobstructed without complicating the build process. * * The two SKUs we’re recommending come with 3x 140mm fans pre-installed at the front. It’s this configuration that allowed Fractal to hit th e top slot in some of our charts back when we reviewed it earlier this year, and even as recently as our Hyte X50 and HAVN BF 360 reviews, the Meshify 3 held at least one of the top entries on the charts. This is a resolute successor to the Fractal Torrent (watch our review), although completely different in the execution of how it got to the top of the chart. * * * * Most of our criticisms of the case were related to its price, but that’s now resolved entirely. We had some smaller complaints, like the cable cutout in the power supply shroud being undersized. Overall though, the Fractal Meshify 3 performs excellently and is a chart-topper in some cases, or top 5 in all cases, while also executing with solid build quality and now competitive pricing against its peers. Fractal’s Meshify 3 Ambience Pro RGB and non-Ambience RGB are deserving winners of our Best Overall Case of 2025 award. The runners-up are HAVN’s BF 360, which engineered the hell out of a unique case, and HYTE’s X50 for its manufacturing feats. Both of these still get separate awards, though. ### **Best Mechanical Design: HAVN BF 360** **HAVN NF 360:Original review | Amazon** The next award is for the best mechanical design. This is given to cases that innovated the most on design execution for complicated mechanical elements, masterful simplifications of complex features, or other well thought-out improvements that might not earn a spot elsewhere in the awards, but should be acknowledged. This year, we’re giving the Best Mechanical Design award to the HAVN BF 360 Flow. Read our review. The BF 360 is a $190 case highly focused on airflow and performance, with extensive engineering behind it. * * * * The case has a lot of smaller mechanical attention to detail: HAVN fine-tuned the bottom ramp for cooling and experimented with different curvatures, bowl shapes, flat ramps, and angles, ultimately settling on what they shipped for its ability to better project air straight into the GPU-side of video card coolers. The case also includes a re-circulation prevention plate (similar to a server rack), so you can plate-out and block the empty area of the front panel after adjusting the fan location. We tested this in A/B benchmarks and found that the backflow prevention plate actually works and does help with thermals. * * In smoke testing, we saw a slight recirculation effect on the inside of the case behind the plate, meaning that this air would likely exit and re-enter without it. Likewise, the panel design uses large slats that are carefully shaped to minimize plastic obstruction at the blade tips of the fans while still maintaining structural integrity of the panel. This is something a lot of companies don’t pay attention to when placing structural support for panels. * * * The BF 360 also has rubber grommets at nearly every metal-to-metal contact point in the case, which we’ve come to appreciate. Panel materials are also of high quality, like the front panel with its faux stone coloring and speckled pattern, which just happens to align with older keyboard and case plastics. * * HAVN also has a number of areas of finer attention to detail and mechanical design, such as carefully-placed cable routing pathways that we almost never found ourselves disagreeing with. These are even labeled and, despite a typo on literally all of them that reads, "Recommendned," we appreciated the thought for newer builders who may need the help. * * * * The case’s fans are a big part of its engineering. Although there is a $160 fan-less version, a lot of what HAVN is doing comes down to the fans. The Flow variant includes 2x 180mm front fans that run 40mm deep, which helps with static pressure performance and gives the fans a wide area of impact. The rear fan is a 120mm x 30mm deep solution. This 3-fan configuration performed well overall, but didn’t quite crack the best rank in our charts. Adding two more 180mm fans to the top situates the BF 360 Flow as the best air-cooled solution we’ve tested yet for our case reviews, though. * * Our review included fan tests on our fan testing machine, including A/B comparisons with the Fractal Torrent for P/Q. This was in addition to our thermal benchmarks for the BF 360 Flow, which had it overall competitive. It’s not the best, but it’s up there with the best and its mechanical design is excellent. Overall, we liked the BF 360 Flow for its design and mechanical complexity. We’d even say we it is “recommendned.” ### **Best Sub-$100 Class Case: Lian Li 207** **Lian Li Lancool 207:Original review | Amazon** Visit our Patreon page to contribute a few dollars toward this website's operation (or consider a direct donation or buying something from our GN Store!) Additionally, when you purchase through links to retailers on our site, we may earn a small affiliate commission. Our next category is for the best case under $100. Despite the lack of competition in the $50-$60 price point in the last 5 years, the sub-$100 point has continued to see competitive releases. If you’re buying closer to $50-$60, we think you should buy used instead if possible. * * * So, for Best Sub-$100 Case, Lian Li holds onto the award for a second year in a row due to its still unbeatable thermal performance and its aggressive, cutthroat pricing. Lian Li continues to show a level of pricing aggression that its competitors struggle to match, at least, until Thermalright finalizes its plans to bring PC cases to the market -- it had promised a $45 case earlier this year which, to Lian Li’s credit, it also mentioned. In the year since the 207 (read our review) won this award last time, it has faced some new competition. We’ll cover those first since this is a repeat winner: * * * * The Fractal Epoch (read our review) case is a Fractal North (watch our review) with a different, cheaper front panel. Otherwise, the tooling is almost entirely the same, the case just loses the panel. Our viewers said Fractal should have called it the South. The price has come down since launch, with the RGB variant including 3x 120mm fans now priced at about $90 on Newegg. That makes it a serious competitor to the 207, but the North was primarily special for its front panel (with the rest of the tooling standard) and the thermal performance of the Epoch can’t beat the 207 in our noise-normalized testing. As for the XT Pro Ultra from Phanteks, we thought the case was simply “fine,” but was otherwise mechanically and thermally unremarkable. The Corsair Frame 4000D RS ARGB (read our review) is a strong competitor, as is Lian Li’s own 217. * * * * * But, of all of these sub-$100 cases that we’ve tested, the 207 is still the best. The 207 includes 2x 120mm bottom-mounted intake fans that pull through the ventilated side panel and the rear of the case, where typically the power supply would sit. These push air straight into the GPU, helping it achieve some of the best cooling on our charts -- even against cases 2x the price. The 207 also has 2x 140mm front-mounted fans and, given the short nature of the case, the air doesn’t have to travel far to hit the CPU cooler and GPU. * * The 207’s noise-normalized CPU thermals put it ahead of the G400A D-RGB, the 4000D RS ARGB, Antec Flux non-Pro (and it’s tied with the Pro), Phanteks XT Pro Ultra that competes on price, and significantly ahead of the Montech XR (read our review) in a similar price class. The noise-normalized GPU thermals also keep it up in the top 3 of our X50 review chart, right behind the Flux Pro (read our review), long-king Fractal Torrent, and within error of the Lancool 217. * * * A lot of the 207’s cooling performance is achieved by moving the power supply to the front of the case, giving room at the back for better-positioned intake fans. The downside to this is the additional complications for power supply installation and cable management, including a sharper bend of the cables and limited depth for packing them all in. In our Lancool 207 review, we talked about how the side panel can bulge without snapping the center snaps if not strictly managing cables, which is resolvable, but requires additional effort and can be annoying -- particularly in cable-heavy builds or with thicker custom cables. * * * * Other than this key downside, the case executes nearly flawlessly on its thermal goals for its price and its price does let it get away with a lot of other tiny annoyances. For example, although it supports larger, we recommend a 150mm-deep power supply as the maximum size to help with some of the cable issues. We still also recommend adding an extra dust filter to the rear former PSU vent, basically the intake vent for the bottom fans, just to help with dust management since it has larger holes than a front panel. * * * * Even with the thermal focus at the price, the case still manages to include a relatively full feature set of velcro straps, USB type-C, ARGB fans, and relatively good build quality. ### **Best Mid-Range Case: Corsair Frame 4000D RS** **Corsair Frame 4000D RS:Original review | Amazon** Our next category is for the Best Mid-Range case. This category is awarded to a case that isn’t too expensive and sticks to the usual playbook for making just a good, mid-level priced case. This year, we’re giving it to the Corsair Frame 4000D RS ARGB. Corsair had a rough few years of case launches, but the Frame series turned it around by reviving the successful 4000D and 5000D cases and overhauling them. * * * * * * The Frame series cases are most interesting for their modularity. This includes a fully removable motherboard tray with a few screws, a power supply shroud that can be easily removed to fully open-up the case, a removable PCIe bracket that can rotate for vertical or horizontal GPU installation, and fan rails with fixed width spacing but free vertical spacing for precise positioning of the fans. * * * * * Corsair has managed to make its Frame case modular without it becoming overly gimmicky, and more importantly, unlike many of its contemporaries, it’s done it without ballooning the cost for a feature most will only use once. Back when we reviewed it, we criticized Corsair for launching its expansible case without any expansions or spare parts. Now, it has a page full of replacement components, extra fan rails, fan clips, glass panels, and so on. Corsair still hasn’t taken full advantage of its plans for alternate panels, though. Its store has spare parts available, which is nice, but the bold vision for alternative panel designs that can just be ordered from Corsair seem to have largely fallen flat. * * * * * * * The Frame 4000D deploys a strange 3D Y-shaped front panel that manages structural rigidity while also having some of the highest percent porosity of any of the cases we tested this year, which helps with airflow. It’s not as clean-looking as the original, flatter 4000D, but does manage to open up the chassis for flow performance without making it so perforated that it becomes flimsy. Corsair also heavily ventilated the top of the Frame 4000D, the rear, and the lower side panel (below the glass). * * With the Frame 4000D RS ARGB including its fans, the case landed in the top quarter of noise-normalized CPU thermal results (between some Fractal North XL listings – read our review), although it was more middling in our GPU thermal results when noise-normalized. For these reasons, the Frame 4000D RS, which is about $90, didn’t get our Best Sub-$100 case. It’s just not strictly better than the Lancool 207 in that price category, but from a build quality standpoint and for its versatility, it’s still a good case that belongs on the list. The ARGB variant that includes fans puts it into the mid-range category at typically about $125, although lately, the case has been available for around $100 with fans. We don’t know if that’s temporary. The Corsair Frame 5000D RS ARGB is basically the same case. But it is scaled up. That one’s closer to $180-$200 with fans. If the Frame 4000D is what you want, but too small, the Frame 5000D fills that segment; however, Corsair does have a ton more competition at the $180-$200 price point. ### **Best Noise-Normalized Case Thermals: Meshify 3, Lancool 217, & Flux Pro** **Fractal Meshify 3:Original review | Amazon Lian Li Lancool 217: Original review | Amazon Antec Flux Pro: Original review | Amazon** Our next award is for Best Noise-Normalized Case Thermals, which combines acoustics and thermals for a simple, objective scoring. This time, there are 3 winners; one for 3 different categories. * * * The winners are the Fractal Meshify 3 Ambience Pro RGB for CPU noise-normalized thermals, the Lancool 217 for VRM & RAM noise-normalized thermals, and the Antec Flux Pro for GPU thermals. For this testing, we position the cases in our hemi-anechoic chamber with a microphone 1 meter away, then adjust fan speeds of the case fans -- but not the system fans -- until it hits a noise-normalized target of 27 dBA. This is using the fans included in the case. This levels the playing field to show some nuance beyond full-speed fan testing. The Fractal Meshify 3 came out earlier this year and immediately netted itself a new best result on our noise-normalized charts. When we reviewed it, the case instantly chart-topped for noise-normalized CPU thermal performance with a 44-degree P-core result over ambient. The Flux Pro, which came out late last year, maintains rank in the top cases for this same chart. Objectively, the Meshify 3 Ambience Pro RGB won this result with its combination of its 3x 140mm Momentum 14 RGB fans and its highly perforated front panel design. In our noise-normalized VRM and DDR5 memory thermal testing, the top performers are the Lancool 217 and BF 360 Flow. Objectively though, across both the VRM and DDR5 thermal tests, the 217 technically is the victor for these. For noise-normalized GPU thermals, the Fractal Torrent is still the top performer, but it’s tied by the Antec Flux Pro (which was one of our winners for thermal categories last year). The Flux Pro is closely trailed by the 207 and 217 Lian Li cases, with the X50 equipped with non-Pro (Flux) fans following those. Considering the Flux Pro’s better performance in other categories than the Torrent (and the age of the Torrent), we’re giving the GPU-side of this award to the Flux Pro for one more year running. ### **Best Out-of-the-Box Case Thermals** **HAVN BF 360 Flow:Original review | Amazon** **Lian Li Lancool 216:Original review | Amazon Lian Li Lancool 207: Original review | Amazon Antec Flux Pro: Original review | Amazon** Next up is the award for Best Out-of-the-Box Thermals, a GN classic. * * * * The winner for Best Out-of-the-Box Thermals is split between the HAVN BF 360 Flow for CPU results, tied with the older Lancool 216, and the Lancool 207 for GPU thermal results, tied with the Antec Flux Pro. With full case fan speeds and testing CPU thermals, the best performer for our data set is the HAVN BF 360 Flow with extra fans installed -- but that’s not out-of-the-box, and this is for out-of-the-box thermals. Stock, the older 216 and the HAVN BF 360 Flow default configuration (without the extra fans) are both at the top of the results. The Hyte X50 with Flux fans is also up there, but that was an arbitrary choice as the case doesn’t include fans, and so it isn’t eligible here because, out of the box, it doesn’t have fans. Like the Torrent, the 216 is old enough that we’re ruling it out; however, it and the BF 360 are at about the same performance level, followed closely by the Lancool 217, which is newer. For GPU thermals, the Lancool 207 and Antec Flux Pro hold the top of the chart. The BF 360 couldn’t quite make it to the top here, with its stock result landing down below the Lancool 217. Antec’s Flux Pro remains highly competitive now a year later, and we look forward to seeing what the Noctua version of the case can do once it comes out. Lian Li’s 207 manages better GPU memory thermal results than the Flux Pro and is tied in GPU temperature. ### **Most Innovative Case: HYTE X50** **HYTE X50:Original review | Amazon** Our Most Innovative Case award is given to cases with the most different designs that take risks. That can be manufacturing, market, or performance risks, because innovation and risk pushes the industry forward. This year, the Most Innovative Case award goes to the HYTE X50 case. We’re giving a runner-up mark to the SilverStone FLP02 as well. * * * * The Hyte X50 makes heavy use of metals, mesh, and curves. The case moves the power supply to its historic top-mount position and aimed to achieve top ranks in cooling performance with its perforations. * * * * * Most notably, the manufacturing process involved figuring out how to edge-wrap fine mesh with moderate porosity to the corners of the panel, past the apex of the curve, and also figure out how to stamp and bend louvered slats to enable unobstructed exhaust while also strengthening the PCIe slot structure. We tested this with Schlieren photography in our lab and found that the louvered vents do actually work. * * The case’s so-called “acoustic” glass didn’t do much for acoustics, although its lamination will help with containing shards in the event it ever shatters. Thermally, Hyte managed to achieve competitive performance when equipped with fans from the Antec Flux non-Pro, including beating the Flux non-Pro with its own fans in the noise-normalized CPU thermal test. It wasn’t the chart-topper that Hyte hoped for, but at least did acceptably. GPU thermals were far more promising for the Hyte X50 Air thanks to its mesh side panel, landing it toward the top of the chart with the Flux non-Pro fans. Our Schlieren photography showed that this was due, in part, to the mesh panel allowing GPU exhaust to exit and get away from the case more easily. The case did OK overall, but that’s not why it gets an award. * * * * The Hyte X50’s notably bubbly design gives it peerless differentiation in at least one category, which is its innovation on the manufacturing process to get there. Compound curves are a complicated manufacturing hurdle when trying to maintain structural integrity with thinner metals and high porosity. Although easily achievable with cheaper plastic molding, doing this with metal meant that Hyte had to iterate through multiple trial-and-error manufacturing processes with its factories. This meant changing metal compositions to modulate strength, which sometimes resulted in rusting, changing the depth of the mesh wrap to increase structural rigidity, and changing with porosity of the holes (such as by reducing porosity at the perimeter to reduce tearing of metal, where the fan’s blades most need a porous surface). Although less challenging, Hyte also brought curved mesh to its side panel for one of the case variations, including mesh all the way up-and-around the top of the case. In our technical discussion with Hyte product director Rob Teller at Computex this year, we learned about the process in-depth. * * * Additional manufacturing maneuvering included using a deep draw stamp for the power supply shroud, a process that’s more time-consuming, but allows more depth to a metal component with curves in it. The louvered slats also required close work with the factory to stamp and twist without ballooning cost or reducing yields. There are areas that it still struggles, such as when our review talked about the creaking or flexing of panels in some locations. The panels can also be finicky when slotting them in place, requiring more attention to avoid an uneven mount. Even with that, we think the Hyte X50 managed to produce a different-looking case without sacrificing too much in terms of performance or other qualities. The $130 to $160 pricing also puts it in range of most of its modern competition, although lacking fans drives the functional price up. #### **Runner-Up: SilverStone FLP02** **SilverStone FLP02:Original coverage** Our fully custom 3D Emblem Glasses celebrate our 15th Anniversary! We hand-assemble these on the East Coast in the US with a metal badge, strong adhesive, and high-quality pint glass. They pair excellently with our 3D 'Debug' Drink Coasters. Purchases keep us **ad-free** and directly support our consumer-focused reviews! * * * * This category’s runner-up is the SilverStone FLP02, a retro-themed case that manages to modernize the functionality of an old look. We appreciated the concept of this case when we saw it at Computex. The case has 5.25” drive bays complete with retro-themed Floppy face covers, although you could complete the look with an actual IDE floppy drive. Center-front is a keyed power toggle, a “turbo” switch that boosts fans to 100% speed, and a reset button, plus a large seven segment display to show the fan speed. * * The FLP02 lacks in some areas, like access to air, but still outdoes its predecessor inspiration while staying true to the style. The case modernizes the sleeper build, like with support for the large video cards of today and some radiator support. For these reasons, we give SilverStone the runner-up award for Most Innovative Case. * * *

The Weirdest Case So Far: HYTE X50 & X50 Air Case Review & Benchmarksjimmy_thang November 24, 2025 ## We test the Hyte X50 and X50 Air for thermals, acoustics, build quality, cable management, radiator and fan compatibility, drive support, and more The Highlights * The Hyte X50 and X50 Air cases take some risks on design but manage to do so while keeping thermal performance overall competitive * We didn’t measure a noticeable difference in noise dampening from the laminated “acoustic glass” * The case is weird, but we like it * Original MSRP: $130-$160 * Release Date: November 2025 #### Table of Contents * AutoTOC Buy a GN 4-Pack of PC-themed 3D Coasters! These high-quality, durable, flexible coasters ship in a pack of 4, each with a fully custom design made by GN's team. You'll get a motherboard-themed coaster with debug display & reset buttons, a SATA SSD with to-scale connectors, RAM sticks, and a GN logo. These fund our web work! Buy here. ### Intro Today we’re reviewing the $130-$160 Hyte X50 airflow-focused case with a bubbly design, and it also has uniquely shaped, louvered vents at the back that should theoretically help evict air from the case faster. At least, that’s what Hyte is claiming. The company told us it was an engineering challenge for Hyte to stamp and retain the steel and then to bend it slightly. _Editor's note: This was originally published on November 2, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication._ * * * #### Credits * * * ##### Test Lead, Host, Writing Steve Burke ##### Testing, Writing Patrick Lathan ##### Camera, Video Editing Vitalii Makhnovets Tim Phetdara ##### Camera, Schlierenmaster Andrew Coleman ##### Writing, Web Editing Jimmy Thang * * * * * * * * We also used our Schlieren imaging setup to illustrate this, which uses a parabolic lens and some lighting tricks to visualize the air density gradient around the case. * Rather than using smoke flow testing with its many inaccuracies and limitations, such as fog density, we can instead look at the refractive index changes caused by air density gradients that distort parallel light beams. A better exhaust path isn’t Hyte’s only claim we’re validating: Hyte says: “Laminated acoustic glass significantly enhances passive noise damping by reducing transmission of sound with a Shatter-Proof(ish) bonding layer.” Typically, laminated glass just keeps the glass together in the event it shatters, which Hyte also says, but the acoustic claim is new to us. So we asked Hyte to send us a non-laminated panel with its laminated panel to test the concept in our acoustic chamber, which produced these frequency spectrum plots that we’ll talk about down below. * * * Mechanically, the case is unique and executed with excellent build quality: The X50 uses a continuous panel around the side and top in either glass or double mesh, with a mesh front panel that has edge-wrapped perforations for cooling. * * Pricing with tariffs in the US is $160 for the X50 with a glass side panel, or $130 for the X50 Air with a mesh side panel. Worldwide pricing is $150 for the X50 and $120 for the X50 Air. Neither case comes with fans, although they do sell color-matched fans. Finally, at least for the review samples, Hyte included incredibly bad smelly candles. ### **Specs** Form Factor| ATX Mid-Tower ---|--- Volume| 63L Case Dimensions| 485mm (H) x 510mm (L) x 255mm (W) Chassis Material(s)| 1mm Thick Steel, ABS Window Material(s) [non-Air]| 4mm Thick Laminated Acoustic Glass Grommet Material(s) [non-Air]| Total Coverage Silicone Motherboard Support| E-ATX (up to 10.6”), ATX, MATX, ITX Power Supply| Up to 223mm GPU Support| Up to 430mm Length, 160mm Height Side Fan Support| 3x 120mm Front Fan Support| 3x 120mm / 140mm Bottom Fan Support| 3x 120mm up to 32mm Thick Rear Fan Support| 1x 120mm Side Radiator Support| Up to 360mm Long & 70mm Thick Front Radiator Support| Up to 360mm Long & 105mm Thick Rear Radiator Support| 120mm Max CPU Cooler Height| 170mm Storage Support| 2x 2.5” Drives, 1x 3.5” Drive Expansion Slots| 7 Horizontal (Full-Size) Front I/O (Audio)| 3.5mm Combo Jack Front I/O (USB)| 1x USB 3.2 Gen 2x2 Type-C, 2x USB 3.2 Gen 1x1 Type A Front I/O (Power)| Tactile Mechanical Power Switch with LED Dust Filters| Bottom, Magnetic Contemporary Feet| [ONLY with these colorways] Snow White, Pitch Black, Wild Cherry Paw Feet| [ONLY with these colorways] Taro Milk, Strawberry Milk, Matcha Milk Warranty| 4 Years Color(s)| Snow White, Pitch Black, [Following colors for non-Air only] Wild Cherry, Taro Milk, Strawberry Milk, Matcha Milk _Specs copied from manufacturer materials, please read review for our own measurements and opinions_ ### **Overview** * * * * The rounded curves of the X50 are its most prominent feature, as well as the hardest to manufacture. The design is promising for airflow, with perforations that extend around the edge of the curve for maximum surface area. Specifically, getting the bent shape with mesh at the edges was difficult to manufacture without using plastic. During a visit to HYTE's office earlier this year, we showed several different panel designs that Hyte had experimented with, all of which were attempting to solve unique manufacturing challenges posed by what they ultimately went with. You should check that article out to see how down-to-the-wire the final concept came. * * * Even as it is, some of the most complex pieces are plastic. Color matching isn't perfect, but on our purple review sample, it's impressively good given that the paint, plastic, and silicone elements all have to be matched, and that almost every piece of the case has color. We have shots of some of the other colors, like glossy red, green, and pink from Computex. Hyte has been experimenting with non-black-and-white colors for a few case iterations now. * * * * * * * The Hyte X50 supports up to 3x 140mm front-mounted fans and has side and bottom support for 3x 120mm fans each. The case also fits one rear 120mm fan, so overall, fan sizes are relatively limited, mostly to 120mm fans. The front mount has 425mm of clearance from top to bottom, so typical 360mm radiators will fit and 420s will not. Radiators up to 105mm thick on the front and 70mm on the side are officially supported, but not simultaneously: assuming 25mm thick fans, you have about 25mm of radiator space on the side and about 35mm on the front before the two conflict with each other. Because the power supply is in the once-traditional top-mounted position, its fan also acts as exhaust. Power supplies are tested in ambient temperatures of 40-50 degrees Celsius, so they can handle this heat better than most components. * * * * The front mount is the only one that's removable. It can be adjusted out in increments of 10mm, starting from either 0mm or 15mm, depending on which way the mounting hardware is flipped. The maximum 45mm offset shown in the manual doesn't actually fit with the front panel. Fortunately, we don’t think that will affect compatibility for anyone. Pushing the bracket back to 0mm means that it's flush with the front of the metal chassis; in other cases, we've found that this can help with thermals, but setting a 0mm offset makes it much harder to reach the cutouts for fan cables on the front of the chassis. * * The X50 Air replaces the laminated tempered glass panel with a cheaper mesh panel. Neither variant includes fans, and all other panels are the same between the models. There isn’t a side mount for fans to take direct advantage of the Air's mesh panel (like the North's side fan bracket, for example), although downdraft CPU coolers and some GPU configurations could benefit from it. Claimed motherboard support goes up to 10.6" for so-called “E-ATX”, but standard 9.6" ATX boards line up precisely with the edge of the motherboard tray and look much neater. Mechanically, the case is mostly a standard box but refined. * * * * * The side panels slide down onto the case and pop into position, with the larger panel that wraps over the top assisted by magnets. The magnets lead to some flexing of the Air's mesh, but the panel is supported on all edges by the shape of the chassis. The X50’s panels need to be pressed against the chassis at the same time as they're pressed down or they won't land flush, which the manual highlights. As we so frequently point out, there's no way to screw down the panels for shipping, although the way the panels are attached makes it unlikely they would ever come off with the case standing vertically. * * * * The power supply shroud is a solid piece of steel and is fully enclosed except the bottom interior. This will help with CPU tower cooler exhaust, but we’d like to have seen an option that could be flipped to take advantage of the top ventilated panel as well. Maybe this was a choice to ensure users of the glass panel don’t accidentally orient the power supply the wrong way, or maybe it was for cost control, but the power supply can only pull from internally. Fortunately, they can handle the warmed air better than most parts. Front-to-back, the shroud has 270mm of space (including cables) with an advertised 223mm maximum PSU length. Good attention to detail means there’s an included velcro tie-down inside the power supply shroud, although less-good attention to detail means difficult to get access to once a power supply is installed. * * The X50 is nearly as wide as the original dual-chamber O11D (at 25.5cm versus 27cm), which leaves about 5.5cm between the motherboard tray and the side of the case for cable management. Typical mid-towers have around 2-3 cm of cable management clearance. * * * * * The hidden paths for fan cables are excellent. The bottom, front, and side mounts each have entrances into the molded plastic channels at the top and bottom of the case, although you may need to use some extensions if you want to route everything exclusively through those paths. The tie points around the edge of the motherboard tray aren't fancy, but they're effective in combination with the included velcro ties. Alternatively, some of the ties can be moved to the back of the case for bundling power and I/O cables, a feature that Fractal has been pushing for a few years now. * * Continuing on cable management: Hyte says the silicone cable grommets are the single most expensive piece of the X50 by volume. These are giant, molded pieces in the style of the Y70. Grommets aren't included with the less expensive Air SKU, with the justification that they can't be seen through the side panel anyway. * * Hyte also has a number of areas of smaller attention to detail in the case. The power button, for example, uses what appears to be a GTMX low profile switch, similar to a Cherry MX Blue keyboard switch. Using a keyboard switch here is 100% on-brand for HYTE, as is the fact that it requires four LEDs just to illuminate the circular white power button. The only downside is that the tactile click doesn't line-up with actuation, so you can click the button without turning your PC on. * * The "paw" and "contemporary" foot styles are also unique. These are locked to specific colorways: so-called paws come with purple, pink, and green cases, and contemporary feet come with white, black, and red cases. These are not available as standalone accessories. * * * The X50's elaborate, rounded drive cage supports 2x 2.5" drives and 1x 3.5" drive simultaneously. The fact that the X50 can fit 2x 2.5" drives and a full-sized hard drive all stacked on top of each other behind the motherboard tray is a testament to the unusually large cable management area and the extra width of the case. Unfortunately, all three drives must be installed with plugs facing the rear of the case, so there's no tidy cable path from the power supply to the drive cage. * * On packaging, there’s some cognitive dissonance on the environmental friendliness: We made a point of praising the BF 360 (read our review) for shipping screws in labeled reusable bags. Hyte also does this, except Hyte’s reusable bags contain smaller disposable bags, which is just totally unnecessary. ### **Thermals & Noise** Grab a GN15 Large Anti-Static Modmat to celebrate our 15th Anniversary and for a high-quality PC building work surface. The Modmat features useful PC building diagrams and is anti-static conductive. Purchases directly fund our work! (or consider a direct donation or a Patreon contribution!) * * * * * * * Since the X50 doesn't come with fans, we gave it the HYTE HS 420 treatment by swapping in a set of fans from a competitive high-airflow case. According to HYTE, "optimal airflow for most configurations are going to involve 3 front intake fans with a single rear exhaust fan." This proved difficult: the Antec Flux Pro (read our review) has more than three 140mm fans, the BF 360 Flow (read our review) has 180mm fans, the Lancool 217 (read our review) has 170s, the Meshify 3 only has three fans total, the Montech XR Wood is from a different price class, and so on and so forth. The compromise we settled on was the Antec Flux non-Pro; although it's a much cheaper case, it comes with four 120mm fans and one 140mm, which all fit simultaneously into the X50. We placed the single 140mm fan in the front top slot as intake, two 120mm fans as intake in the other two front slots, another 120mm fan as intake in the side bottom slot, and the single reverse-blade 120mm fan as rear exhaust. The bottom slots were left empty. The HAVN BF 360 is the head-to-head competitor for the X50 in some ways as they are both thermally targeted, followed by maybe the Antec C8 Curve (we have results for the C8 ARGB, which includes fans). Of the cases that HYTE chose to compare the X50 against in reviewer materials, we agree most with Fractal's Torrent (watch our review) and North XL (read our review) (and the NZXT H7 Flow 2024, but we don't have test data for that one). #### **Acoustic Glass Testing** To start, we did some tests in the hemi-anechoic chamber with our microphone positioned one meter away from the case. ##### **Frequency Spectrum Comparison** An important note first: Our standardized acoustic testing for cases measures straight-on at the front panel, which we did take for our noise-normalized process and for which the X50 ran at around 44-45 dBA under those conditions. For the panel comparisons though, we’re shooting straight on at the side panels instead since glass on the side won’t affect front-facing noise much other than emitting more of it forward than a mesh panel. For the frequency spectrum comparison, the laminated acoustic panel is nearly identical to the unlaminated panel. On a technicality, the laminated panel shifts some of its peaks slightly left on the frequency plot, such as at about 220-230 Hz, about 500 Hz, and about 900 Hz. They spike at the same points, but the laminated panel is very slightly lower frequency. It is hardly measurable and difficult to know if this is variance or not. We don’t think it’s variance, though. The noise levels are the same: The laminated panel measured at 40.2 dBA and the unlaminated panel measured at 40.3 dBA. 0.1 dBA is absolutely within variance and error, and so these are effectively identical. Adding the Mesh panel version to the charts, we see a 40.5 dBA result, so only slightly higher. The frequency shifts substantially though, despite only a 0.2 dBA increase in dBA SPL. The mesh panel has louder high frequencies on average, especially after the 1000 Hz marker. The range of 1,500 to 5,000 Hz in particular increases. Also notably, the mesh panel sees a dip in noise levels for the frequencies in the range of 250 Hz to 500 Hz. The same spikes still exist, with the 230 Hz or so spike present but quieter, while the 450-500 Hz spike has shifted slightly right on the frequency axis. * * We certainly believe that so-called “acoustic glass,” or laminated glass, can do something, but in this context, it seems like opportunistic marketing that retrofits itself into the actual purpose of the lamination, which is simply that the panel has to be made that way by the factory. It doesn't hurt, but the front of the case is mesh with a bunch of holes in it, and the side panel doesn't change that. #### **CPU Full Load Thermals - Noise-Normalized** For noise-normalized testing, we lower the speeds of all case fans until the total noise level measured at one meter in our hemi-anechoic chamber drops to 27 dBA SPL. This required very slightly different fan speeds for the X50 and X50 Air, but that’s exactly why we have this chamber. Room-level measurements would have had us at about the same noise level, which would slightly affect the results. Adjusting for this, the CPU thermals in the two cases are identical at 43 degrees Celsius above ambient for all-core and 46-47 degrees for the P-cores. These results were gathered with the Antec Flux's stock fans. The BF 360 had a strong showing in our recent review, and it's ahead of the X50 here with a significant lead from its 41-degree all-core average. The once-king Torrent is basically tied with the X50, as is the mesh-sided North XL. Appropriately, the Antec Flux is only a little warmer than the X50 (using exactly the same fans) at 44 degrees, tied with the glass-sided North XL. Other than margin of error, the other differences would emerge from the case design. The Antec C8 with its bottom intake averages significantly worse for CPU thermals at 50 degrees. #### **GPU Full Load Thermals - Noise-Normalized** In the same test, the X50 and X50 Air are more differentiated when considering GPU thermals. The X50 averages 44 degrees Celsius above ambient on the GPU and the X50 Air averages 42, which is a relatively large gap for GPU thermal testing in our suite, especially for the same frame and fans. This makes sense, since the mesh panel gives the GPU external access to air and its exhaust from the top of the video card can also easily exit the side panel. The BF 360's performance is strong in this test at 42.5 degrees, as is the C8 ARGB due to the bottom intake, but the X50 Air ties both for GPU thermals. The Torrent remains best on the chart with a 40 degree average, while the better of the two North XL results ties the X50 non-Air, and the Flux falls behind at 46 degrees. #### **CPU Full Load Thermals - Full Speed** Moving to tests with the case fans at max speed, the best CPU temperature average was scored with the front bracket in the 0mm offset position, with a 36 degree all-core average and 39 on the P-cores. That's followed by the regular X50 at 37 degrees all-core, and then finally the X50 Air at 39 degrees all-core. For CPU thermals, it makes sense that the glass panel would help as it ensures all intake makes it to the CPU (rather than having some of it escape out of the side panel prior to hitting the cooler). This is getting complicated, so we'll focus on the regular X50 for comparison here. The BF 360 Flow ended up at 40.5 degrees for P-core thermals when running at 44 dBA, with the X50 at the same noise level -- so we’re basically noise-normalized again -- and tying it. The X50 Air falls behind here at 42.8 degrees since we’re losing CPU-bound air through the side panel. The Torrent's 42-degree average at 42.3 dBA is still respectable, but the flood of high airflow cases in recent years has pushed it down the chart, where it sits alongside the Antec Flux. The Flux being at 41.9 dBA and 42.7 degrees shows that its case kills more of the noise than HYTE’s X50 (which is about 3 dBA louder) since they have the same fans here, though the X50 gets more air into the case. The C8's CPU thermals continue to be relatively weak with a 46 degree all-core average, although that's accompanied by a relatively quiet 37.1 dBA noise reading. #### **GPU Full Load Thermals - Full Speed** GPU thermals aren't significantly affected by the change in side panel between X50 and X50 Air, with both averaging between 42 and 43 degrees above ambient for GPU temperature. This is a reversal of the behavior we saw in noise-normalized testing, so there's value to considering both. With a 0mm offset on the front bracket, that average rises to 44 degrees, so it's not an upgrade across the board. The stock BF 360 Flow averages 40 degrees, which beats the best of the X50's results. Overall, the BF 360 has either tied or beaten the X50 in this test. This obviously depends on what fans are installed in the X50 and where, but we were fairly generous to the X50 by installing five total fans from the Flux (non-Pro). #### **CPU Full Load Thermals - Standardized Fans** As usual with cases that don't ship with any fans, our standardized fan testing assumes greater importance. Of our standard set of three Noctua fans, we installed the 2x 140mm fans as front intake and the 1x 120mm as rear exhaust, which comes close to the 3x intake/1x exhaust configuration that HYTE considers optimal. With the 2x 140mm fans positioned in the middle of the front panel, the X50 and X50 Air performed identically, at 41 degrees all-core and about 44 on the P-cores. Shifting the fans down to the lowest front slots in the X50 slightly lowered CPU temperatures. The BF 360 non-Flow comes without fans, so this test is perfect for comparing that variant against the X50. Using the same set of fans, the BF 360 averaged 41 degrees all-core and 45 on the P-cores, which ties it with the X50 when the standardized fans are positioned in the middle, but the X50 pulls ahead a little with the lower fan placement. For context, when we performed this test in the BF 360, the intake fans were positioned as low as possible on the front panel to point directly into the airflow scoop. Generally, this test indicates that the CPU cooling performance is equal between the BF 360 and X50 with our setup. #### **GPU Full Load Thermals - Standardized Fans** Moving to GPU thermals with the standardized fans, the X50 and X50 Air with the front fans in the middle position again performed identically at 47 degrees above ambient, and the lower position was again better at 46 degrees. The BF 360 had particularly good performance here, likely due to the airflow scoop, averaging 41 degrees above ambient. Combined with the CPU cooling results, this puts the BF 360 ahead of the X50 when using the same set of fans. ### **Schlieren Imaging Air Density Test** * * This Schlieren imaging air density test example shows the case without any side panel installed at all. You're just seeing the true chaos of air entering and exiting through a wide open side panel where there's no particular direction of flow. * * Adding the glass panel has the expected effect. The only air density gradient we see now is from air exiting the back of the case and floating around near the mirror. There's not much movement here, as expected, because air can't penetrate glass. Swapping the panel to mesh instead, we see something pretty cool. The gradient change is most visible exactly where the GPU's exhaust is. It's spewing air out of the mesh from the GPU, allowing it to exit the case more easily. And you can also see it going pretty far away from the case, which is a good thing. This has some benefits, but because the glass had provided a sort of railroaded path for intake to hit the CPU tower, it also has some downsides. We can also see, for example, towards the bottom that some air is exiting before ever hitting the GPU or the CPU tower, meaning that air is taken in from the front bottom position and effectively lost without doing anything actually useful in this configuration. In this shot, you can see how the air, or specifically the air density gradient, although still chaotic on exit, takes a relatively straight path out of the louvered vents. * * The vents nearest the blade tip of the fan show this the most where we can actually see how the air deflects around the vent punch out. That's the white sort of tip around the edge of that. * * * This shot shows us the vents as well where we get straight clear paths and gradient density changes prominently seen right at the vent tip. Once again, the air density differences show the air naturally finding itself going up as we'd expect. It’s pretty cool to visualize this stuff, especially that mesh side panel and what it's doing. ### **Conclusion** Visit our Patreon page to contribute a few dollars toward this website's operation (or consider a direct donation or buying something from our GN Store!) Additionally, when you purchase through links to retailers on our site, we may earn a small affiliate commission. The HYTE X50 overall has sturdy build quality and creative design elements, but it also has creative marketing. * * Sometimes, like with the louvered vents, the marketed feature actually does something. Other times, like with the so-called “acoustic glass,” it’s not something we’re able to quantify in our testing. The challenge with testing came down to fans, since the X50 doesn’t include any. By using the Antec Flux fans, we were able to see that it’s generally at least better than the Antec Flux. The standardized fan test is also overall favorable for the X50. To be very clear, there’s an Antec Flux Pro model as well, and that one sorts out differently as it has six fans including 140mm ones. As we mentioned in our BF 360 review, HAVN is HYTE's biggest competition for this release: the fanless BF 360 costs $160, just like the X50 (in the US). HYTE is well positioned, though, because if you want a case that looks like the X50, there isn't any competition from anyone. Until GameMAX and Gamdias rip it off like they do with everything else. The BF 360 Flow wins in terms of thermals, it optionally comes with stock fans, and it has all the other benefits of build quality and ease of use that we mentioned in our review, but it's a more conservative design and only available in black and white. We like the X50 and we're comfortable adding it to our list of recommendations alongside the BF 360. As usual, we have yet to see other manufacturers beat Lian Li's price-to-performance with the Lancool 217 at $120 and 207 at $90. But that's not the market HYTE is aiming for. If the colorful design of the X50 is more appealing than its thermal performance, the best alternatives we can give are Thermaltake's Tower 600, Ceres 350, and View 600, although we haven't reviewed those models. * * * As for the rest of the market at the crowded $200 price point (factoring in the cost of fans), even with the higher US prices, the X50 compares favorably to established cases like the C8 Curve, Torrent, and North XL. It does so by combining high airflow with competent cable management and a quality of design that goes deeper than the outside of the case. It's not that the X50 is inherently better than the North XL, for example, but it can match performance in a way that makes it a valid alternative without sacrificing functionality. That applies to the cases we brought up for comparison to the BF 360, as well. Cases like the Meshify 3 Ambience Pro RGB (read our review), Frame 5000D, and Flux Pro. The key factor is the way the X50 looks. The painful back-and-forth with HYTE's factories to get a working process for creating the curved front panels and louvred vents paid off, and it has put HYTE ahead of its rivals. The glass lamination doesn't really matter unless you smash your case, but the big curved panel is impressive even ignoring that, and we actually prefer the glass X50 to the X50 Air. The X50 Air isn't equipped to take advantage of its mesh side panel, so with most setups there's no reason to expect a difference in thermal performance between the models. In our opinion, the curved glass and variety of colors available with the normal X50 make it worth the $30 upcharge. We can't think of many case companies that have been offering this kind of variety other than Thermaltake, and maybe Montech or Geometric Future. We look forward to HYTE expanding the X-series lineup in the same way as the Y-series, hopefully with more sizes and design collaborations. * * *

Valve Steam Machine, Desktop SteamOS, Steam Frame VR, & Controller | ft. Engineering Discussionjimmy_thang November 14, 2025 ## We run through the details of Valve's new hardware (Steam Machine, Steam Frame, and Steam Controller) and interview many of the company's engineers working on the products The Highlights * Valve’s Steam Machine is a mini-PC that runs SteamOS * The Steam Frame VR headset also runs on SteamOS, but is equipped with an ARM-based SOC that will be able to emulate x86 games * The new Steam Controller makes some major iterations on the original, with a particular focus on TMR thumbsticks and connectivity #### Table of Contents * AutoTOC Grab a GN Tear-Down Toolkit to support our AD-FREE reviews and IN-DEPTH testing while also getting a high-quality, **highly portable 10-piece toolkit** that was custom designed for use with video cards for repasting and water block installation. Includes a portable roll bag, hook hangers for pegboards, a storage compartment, and instructional GPU disassembly cards. ### Intro Valve had several hardware announcements today. _Editor's note: This was originally published on November 12, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication._ * * * #### Credits * * * ##### Host, Writing, Editing Steve Burke ##### Writing, Research Patrick Lathan ##### Camera, Video Editing Vitalii Makhnovets ##### Video Editing Tim Phetdara ##### Writing, Web Editing Jimmy Thang * * * * * * * * * The company is announcing its Steam Machine mini-PC with SteamOS, the Steam Frame VR headset with separable compute portion, accompanying Steam Frame VR controllers, and the new Steam Controller. There are multiple aspects of software wrapped into these, like SteamOS on ARM, FEX EMU, Proton updates, and, of course, SteamOS for desktop. At our own expense, we bought plane tickets and booked a hotel, then flew out to Valve’s headquarters for engineering interviews -- and we got some A-tier information from the thermal, acoustic, and electrical engineering teams. The release date for all of these is 2026 sometime, with at least the Frame targeting early 2026. Pricing is TBD. We were told that the Steam Machine will be priced as “an entry-level computer,” but without a range on what that means (and everyone’s interpretation is probably a little different). Valve did emphasize that it won’t be priced like a console, though, so we’d anticipate higher than console prices. There’s no telling where RAM prices will be when they launch. Valve told us that the Frame won’t be priced higher than the Index, but was non-specific on which SKU or bundle. This article will focus on the news and engineering interviews for each component. We have at least two separate videos coming up that we split-out just because we have a ton of depth for all of them and we didn’t want to chop it down. Our next video covers the Steam Frame acoustics, power, and thermal engineering. We will also have a video about the Steam Deck with one of its engineers. For now, let’s get into the specs and news for SteamOS (which is a Linux-based operating system) finally coming to desktop, the Steam Machine, the controller, and the controllers for VR. We’ll be peppering in relevant interview quotes we conducted with Valve throughout the article below. Valve has attempted to launch Steam Machines in the past, but never committed to a rollout. This time, they’re serious about it and they’re shipping a first-party solution. This also means that SteamOS is getting an official shipment to desktop, which should bolster support for self installs even if Valve isn’t fully tuning it outside of the Steam Machine. Let’s start with the Steam Machine. ### **Steam Machine** We'll start with the new Steam Machine, which Valve claims has "six times" the performance of the Steam Deck. #### **Specs Summary** Starting with the specs: * * * * * * * The Steam Machine uses a BGA solution with SO-DIMM memory and modular M.2 storage, shipping in 512GB and 2TB models (with a microSD slot for additional storage) and optional Steam Controller bundle. MicroSD cards can be pulled relatively seamlessly from the Steam Deck and Steam Machine and transplanted to each other with games on them (in an easier way than Windows). The CPU is an AMD 6-core, 12-thread Zen 4 processor with its IGP fused off and the GPU is a separate "semi-custom" RDNA3 28-CU chip that supports ray tracing, paired with 8GB of GDDR6 VRAM. RDNA3 was the RX 7000 series, with the 32-CU RX 7600 being relatively close as a CU-count comparison, though there are other specs to consider as well. System memory is separate at 16GB of DDR5-5600 laptop memory. * * * For rear ports, the Steam Machine has one DisplayPort 1.4, one HDMI 2.0 (with CEC), one 1Gbps ethernet, one USB Type-C 3.2 Gen 2 (10Gbps), and two USB 2 Type-A ports. There are also two USB 3 Type-A ports in the front, 2x2 WiFi 6E, and another two separate antennas, one for Bluetooth and one specifically for Steam Controllers. Out of the box, it ships with SteamOS. * * * * * Our biggest immediate concern without having tested it is how closed-off the box is in general, with the single 120mm fan used as a pull solution through the bottom, which has minimal elevation, and holes in the front. The front plate pulls to provide much greater access to air. We can’t really pass too much judgement here until we have it to benchmark, which we will definitely do. Valve did a lot of tuning on thermals. The system uses a 120mm fan that’s used as a pull solution, which will pull air through the bottom and the holes in the front. #### **Details** There were two critical shortcomings with the original Steam Machine project: first, Valve never shipped it, although we did cover third-party small form factor PCs intended to become Steam Machines. More importantly, Valve never planned any first-party production (beyond a small beta run) and there was no reason to use the 2015 version of SteamOS over Windows. **Valve Engineer Yazan Aldehayyat:** "So, essentially, we just feel like now is the right time for us to make this happen, just because we finally feel like we have the software and the hardware pieces all figured out. By all accounts, the SteamOS experience on Steam Deck, people love it, they love the convenience. We're seeing a lot of people use their Steam Deck docked to their TV. Primarily to us that just means that console, you know, experience is there. But we just simply feel like if we were to make something that's more, you know, intended for that use case, that doesn't have the weight and battery life limitations, we can give those people a much better experience." Although shaped like a console, the Steam Machine is just a very small computer: there's a usable mouse-and-keyboard desktop mode despite the clear living room focus. The BIOS is accessible and modifiable (and navigable with a Steam Controller). From what Valve says, it sounds like it'll be a limited laptop-style BIOS, but Valve mentioned users will be able to "play around with frequencies." We asked Valve about its choice to go with a BGA Zen 4 solution rather than something socketable, seeing as that’d benefit from AM5’s likely immortality if it’s anything like AM4. Valve considered modular CPUs and GPUs, but settled on soldering both chips to the motherboard for thermal and size reasons. Unfortunately, this means it’s not upgradeable, but it gives Valve more control to make it smaller. The CPU has a peak frequency of 4.8GHz, which sounds similar to an R5 7640U and falls short of something like an R5 7600. The silicon is off the shelf, but has been customized with firmware, software, and fusing off certain elements. Valve told us any included IGP is fused off, meaning AMD semi customizes these. Valve told us the Steam Machine doesn’t have a discrete chipset and instead routes all I/O through the CPU. The discrete semi-custom 28CU GPU has a maximum sustained frequency of 2.45GHz, although its peak opportunistic boost would be higher. **Valve Engineer Yazan Aldehayyat:** “I mean I was on the motherboard, so I'm really proud of how small it is for what it is, it's, you know, both the CPU and the GPU's on it. We use the same thermal module to cool everything, so both the CPU and the GPU, the memory, and the FETs as well, are all connected to the same thermal module.” **Gamers Nexus** : “Do you know, off the top of your head, the sort of power consumption during peak load, or nominal load?” **Valve Engineer Yazan Aldehayyat:** “So the GPU is anywhere between 110-130W, we're still dialing that in to be honest. CPU is closer to 30W. And, you know, once you add in memory and peripherals and all that kind of stuff, we're typically in the 200+ Watt range." In a later discussion, Valve told us it locked the GPU in at 110W TDP. We asked about the VRM, although we'll have to wait for a teardown to get final answers, but this is what Aldehayyat said off the top of his head: **Valve Engineer Yazan Aldehayyat:** "Six phase for the GPU, and then for the CPU we have two or three phases, I don't remember exactly. The problem is, like, there's different rails, and so, like, the GPU there's, like, six phase for core, two or three phases for memory. For the CPU, it's, I think, total about four phases. Split between different rails." Power is handled via 12V only. On a pure technicality, it’s not necessarily 12VO in the sense that it is using a blade connector instead of going through cabling. This helps with things like space. The system also uses a 300W PSU that’s mated directly to the motherboard with a blade connector that Valve says hits over 92% efficiency in "most scenarios." Mini PCs frequently cheat the volume calculation by using external power bricks, so it's promising to see Valve commit to taking up so much of the Steam Machine's volume with a true internal PSU, especially given the thermal challenges. **Valve Engineer Yazan Aldehayyat:** “We spent a lot of time thinking about airflow. Living room is actually a really challenging thermal environment, right, like people can shove them in corners, put them behind things, so we thought about how do we have intake and exhaust that is hard to block for somebody who's not really thinking about it. * * Like, so we have bottom intake, so airflow kind of just flows around the front here, through here. And then we also have perimeter intake here, 360 again, because it's really hard to plug every single one of those. And then finally we have exhaust on the back because if you have cables attached here, you can't really push it all the way up against the wall, so you're always guaranteed some space." * * The legs are short, which we pointed out, but Valve said its goal was to get the Machine to fit under a normal 6-inch shelf in a media unit, with the design team fighting for fractions of a millimeter. Valve claimed that the leg height allows for around 95% of the flow performance as a fully unrestricted bottom intake, which we’ll test as soon as we get it in. You absolutely need to make sure this does not go on carpet that conforms to the bottom, though. * * * * * * Valve had some interesting solutions to its airflow and size limitations: There's a plastic ridge on the bottom of the Machine intended to block hot exhaust from recirculating underneath by dropping down slightly to the floor. Under normal conditions, we were told that the airflow split between the bottom vents and the perimeter vents should be 50/50. The rear of the case is made of metal for the highest possible porosity while remaining sturdy and the front is a magnetic plastic panel with the intention that users can print their own, maybe with better airflow. Valve showed us a number of alternative panels, such as an e-Ink display that they won’t be selling, but that they used for testing. * * * The relatively large aluminum finstack is mostly a single, shared solution: It contacts the CPU with phase change TIM, uses paste for the GPU, and then thermal putty for the rest. It doesn't directly contact the system memory or storage because neither are soldered to the motherboard. * * * * * * The heatsink has four heatpipes that look like roughly 8mm solutions, paired with a single custom 120mm exhaust fan at the rear of the box using a custom cowling that gives slightly more depth in some areas for better pressure performance. Using an axial fan rather than a blower for a device in this form factor is unusual and created design difficulties, as opposed to (for example) the PS5, but was necessary in order to get enough airflow. The fan is heavily focused on static pressure, ranges between 500-2,000RPM, and has a blade-to-frame clearance of somewhere between 0.5-1mm. Valve wasn’t certain off-hand if the material is PBT or LCP. **Valve Engineer Yazan Aldehayyat:** "One of the questions that people ask is like, how did you arrive at this architecture? And the answer, I think, is probably going to be surprising to most people, which is we literally designed the fan first. Because like, when you… the laws of thermodynamics are not flexible, right? If you knew how much heat you're trying to remove, you can—and you know the temperature the GPU's trying to reach, that tells you how much airflow you need. If you know how much airflow you need, you can kind of get to design the fan. So we really started by looking at what's on the market, how much airflow they can get at what static pressure, playing with blade geometries, like getting the optimal fan that we can." The memory is upgradeable with disassembly of the Machine. We’re not sure if it’s two sticks or one at the moment. Replacing the stock M.2 2230 SSD should be significantly easier, and there's enough room for a standard 2280 drive. The GPU's GDDR6 VRAM presented its own challenges, leading to a ten-layer PCB design. **Valve Engineer Yazan Aldehayyat:** "Adding layers is typically done because you need the layers for routing, and that for us was the main reason, because we needed to… GDDR6 is really hard to route.” **Gamers Nexus:** “Why is that?” **Valve Engineer Yazan Aldehayyat:** “It's a lot of signals, very high speed signals, you know, they're very dense, so it takes a lot of layers to do that. And also, DDR's pretty power hungry, so you have to kind of use some layers for power routing and then other layers for signal routing.” **Gamers Nexus:** “And is it still true that each trace for memory needs to be the same length?” **Valve Engineer Yazan Aldehayyat:** “Yes. So we have to time delay, it's not even the same length, it's time delay. So like, keep in mind the propagation delay, because the top layers will have slightly different propagation delay than the middle layers, so you have to keep that in mind. Impedance, as well. So we have, basically all the bits have to arrive at the same time. The speed of light, you know, down to, I think, like a picosecond, if I remember correctly. [...]GDDR6 is really challenging. So, like, we had to do things like back drilling, I don't know if you're familiar with that." **Gamers Nexus:**   “I don't think so.” **Valve Engineer Yazan Aldehayyat:** “It's basically like, you have a hole through the board to transition through layers, just a through hole. So if you're trying to go from like layer one to layer five, you have a stub that goes all the way down to layer ten. And what happens then, when the signal transitions through, it hits that, and some reflections go through the via, and they resonate and cause interference.” **Gamers Nexus:** “Is this the same as a via? Or is it different from a via? Because vias I'm familiar with.” **Valve Engineer Yazan Aldehayyat:** “It's basically, if you think about like the path that the signals take, right, they go from layer one, and then go to layer five, and they continue on layer five. Because the via has to go through the whole board, there's a path where it doesn't take. Like a stub, like it's a junction of the road, and the signal wants to take one way, but there's the other one which is a byproduct of just the manufacturing technology. So we have to go back and drill that path and remove it, so the signal doesn't travel both paths and cause interference. **Valve’s Jeff Mucha:** “Backdrilling removes part of the via to only the part you're using." * * Finally, there's a customizable LED bar on the front of the Machine that's used as a status indicator for when the display is off: for example, it can show download progress of games. We're well-equipped to test all of Valve's thermal, airflow, acoustic, and power claims, so look forward to that at launch. ### **Steam Frame** Grab a GN15 Large Anti-Static Modmat to celebrate our 15th Anniversary and for a high-quality PC building work surface. The Modmat features useful PC building diagrams and is anti-static conductive. Purchases directly fund our work! (or consider a direct donation or a Patreon contribution!) * * The Valve Steam Frame’s key feature is a bandwidth-saving foveated streaming solution, which is similar in some concepts to foveated rendering. The Frame is built modularly with a separable, fully contained face-mounted Arm architecture computer running SteamOS. We last covered VR in detail back during the launch of the Vive and the Rift, back when the industry was going crazy slapping "VR Ready" stickers everywhere. That was a crazy time. It’s good that we as an industry have moved beyond two-letter initials for marketing...oh, wait. Meta’s own ARM-based solutions are some of the more popular right now and Valve has launched the Index in the meantime, which it says is the price ceiling for the Frame. #### **Specs Summary** * * * * * The 440g headset consists of a front-mounted 185g core unit and a back-of-head-mounted battery. Valve uses an onboard Snapdragon 8650 ARM solution with 16GB LPDDR5X of unified system memory and VRAM. The headset will have models with 256GB or 1TB of storage. The Frame has a microSD card slot, pancake lenses, 2160x2160-resolution LCDs, a 120Hz refresh rate with 144Hz “experimental mode,” and FOV up to 110 degrees. The 4 exterior cameras are monochrome but IR sensitive and the headset itself contains exterior IR illumination. Valve uses two internal eye tracking cameras for its new foveated streaming feature that we’ll get into soon. * * * Audio deploys four speakers built into the headstrap and two microphones in the headset. Valve also has a PCIe Gen4 custom connector for future accessories or attachments, potentially for third parties, with an example given of support for a color camera. The headset can also stream games directly from a PC using a 6GHz included USB router, bypassing traditional wireless routers and communicating straight between the PC and Frame. * * The controllers have capacitive sensing for finger tracking, TMR thumbsticks (identical to the new Steam Controller's), a full set of standard controller buttons, regular AA batteries, and "Full 6-DOF tracking and IMU support." #### **Details** The Frame’s SOC is a Snapdragon 8 Gen 3 8650 on a 4nm process node. An ARM SOC in a headset isn’t new, but SteamOS on ARM is. This includes ARM-native compilations of DXVK and VKD3D, and it’s also the first time SteamOS is meaningfully expanding out of handhelds and into VR. As it did with the Proton translation layer for SteamOS software compatibility, Valve has devoted significant resources to the development of FEX emulation. **Valve Designer Lawrence Yang:** “Similarly to how for Steam Deck we had to make a lot of advances in Proton to make sure that was really solid before Steam Deck could be an actual viable product, for Steam Frame we did a lot of work on FEX, which we see as a… " **Gamers Nexus:** “Is it a translation layer?” **Valve Designer Lawrence Yang:** "It's actually an emulation layer, so it's emulating the instruction set from x86 to ARM, so it's actually doing both Proton and FEX so that you can play x86 games standalone." Like the existing Deck Verified program, there’ll be a separate Steam Frame Verified program specifically for running games locally on the Frame. **Valve Designer Lawrence Yang:** "In a similar way where when we launched Steam Deck we had a lot of work to do on Proton, and as time went on we patched a lot of the holes, and then it became more of a surprise when a game _didn't_ work through Proton on Steam Deck. We're doing the same process with FEX as well. So like, right now we're pleasantly surprised that so many games are just working, and we know that once we get these into developer hands, we'll be able to learn a lot more and ahead of launch we'll have a sizable library." Valve also told us that developers will be able to ship VR APKs on Steam, and for games with multiple versions, Steam will automatically choose the appropriate build (but it can be overridden). Valve stated "APKs can also be side-loadable just like any non-Steam applications on Steam Deck. We expect that VR APKs that don't leverage proprietary APIs to just work." The microSD card slot should theoretically allow seamlessly swapping libraries between other SteamOS devices on a single card, such as the Steam Deck and Steam Machine. Valve confirmed this for the Deck-to-Machine already. As for use cases, Valve is pushing for the "use your headset as your display" approach, even for non-VR content. Thus far, nobody (including Apple) has been able to fully sell a mainstream audience on this approach. **Valve Designer Lawrence Yang:** "So, I would say Steam Frame kind of represents a fundamental shift in the way that we think about VR: rather than being a PCVR accessory for your PCVR titles, we see Steam Frame as just a new way to experience your entire Steam library, whether it's VR or non-VR. And that high level decision kind of filtered down into all of the product level decisions that we made for the product." Valve seemed primarily focused on streaming from PC-to-headset for the rest, which makes sense since Valve’s biggest advantage over Meta is years’ worth of games that were developed to be run on actual gaming PCs. **Valve Engineer Jeremy Selan:** "Every single device comes with a dedicated wireless adapter. This is USB 3, and it's a 6GHz native router, and it talks directly between your PC and this device. [...] So there’s a whole bunch of channels and that allows us high bandwidth, high robustness, and low latency, and it’s a dedicated link just between these 2 devices. Similarly, this headset has two radios and antennas, which means we can guarantee that one of those antennas is locked off to this dedicated connection on 6GHz while WiFi is talking on the 5GHz spectrum simultaneously to your local router." The dedicated wireless dongle isn't a new idea, but Valve has planned around it from the ground up. It's not strictly required, though; Valve told us that "The Steam Frame headset is Wi-Fi 7 capable, and when paired to your existing Wi-Fi, other streaming features (including Foveated Streaming and simultaneous 5/6Ghz ) are available and work great." Alongside the dedicated wireless link, the second piece of Valve's pitch for local streaming is "foveated streaming." The naming is inspired from foveated rendering, which is a related concept that uses eye tracking to deprioritize the screen area in your periphery and render it at a lower resolution, allowing increased framerate. High FPS remains critical for VR headsets to avoid motion sickness in users, and so these technologies are used to focus the “density” of image quality on areas the user is looking at. * * * Eye tracking updates every 8-12ms according to Valve, although the hardware could go faster. This meets Valve's goal for stream quality while reducing power usage. * * * When we looked at Valve’s debug tool, the company shows user eye position and reflects the area of highest-quality rendering based on that. We tested this for a minute or two at the end of the session and couldn’t immediately notice any lag in the render quality shift, although we haven’t had time to fully analyze it. The software also showed where in the frame lifecycle and rendering pipeline the most time was allocated. It harkens back to some of our talk about simulation time error, aka animation error. Here’s what Valve said about foveated streaming: **Valve Engineer Jeremy Selan:** "So what we're doing is at the streaming layer, we're calling it foveated streaming, where the areas that you're looking, we send essentially the full, high quality content at the highest fidelity we can. And then for everywhere else you're not looking, we're borrowing those bits and also throwing them at that place where you're looking. So for instance if we had, say, one-tenth of the pixel area is where you're looking, which is a very common number, then that's effectively a 10x multiplier in effective bandwidth, quality, as well as reduced latency, because we can steer all of those bits to where you're looking. And that only works because the eye tracking is fast enough that as you look around the scene, we send the bits to where you're going to look before you land there." If foveated rendering weren't already an established technology, we'd have a hard time believing that foveated streaming could be so reactive, but in the brief testing that we did with it, we couldn’t see any changes or difference in image quality. There are still a ton of variables involved (like the performance of the host PC), but when we asked Valve for a ballpark end-to-end system latency when streaming, we were told 10-20ms was the target, which includes encoding, transmission, and decoding. * * Target bandwidth is 250Mbps, regardless of whether the headset is running wireless or wired, so that's the baseline for that loose 10x multiplier in effective quality that Valve is claiming. Everyone seems pretty confused about the Quest 3's max bandwidth, so part of Valve's advantage here is that it's shipping the streaming hardware with the headset, so everyone gets a consistent high-performance experience: for streaming, it doesn't matter what router you have or whether your PC has a wired internet connection. Foveated streaming is a system-level feature, so no additional support is required from developers. * * In contrast to previous Valve headsets, the Frame is a standalone battery-powered unit with inside-out tracking: no base stations and no mandatory wires. It ships with a 21.6Wh lithium ion battery. Valve also informed us that the Steam Frame hardware is not cross-compatible with the Index and Vive at this time. Valve claims that power consumption from nominal to peak loads can range from approximately 7W (when streaming) to more than 20W (when running games on the headset itself). Power is heat, so at 20W, Valve’s thermal solution comes more into play. * * * * * The headset is separated into two modular units: the core and the headstrap. The face gasket is attached with magnets, so the grossest part of the headset should be fairly easy to replace, while the core is fastened to the headset with three toolless clips. The core contains the optics, the compute, and cameras, while the headstrap contains the audio and battery. And because everyone with a 3D printer is looking for a problem to solve, we’re sure people will make quick use of the modularity in this design. We have a huge amount of detail on the Frame's thermal solution and how it keeps components (and users) cool, but we'll save most of that for an immediate follow-up video to this one. * * * * The short version for today is that there are two heatpipes flanking the silicon, with the PCB area at the backside of the BGA package cleared to allow rear heatpipe contact and conduction through the PCB. Both of these heatpipes route to a finstack at the top of the headset, near the browline, with ventilation pointed up, out, and away from the user. There's a single custom blower fan in the core module that's governed by thermistors across the PCB in addition to chip sensors. The displays are hotspots, but that's not necessarily a problem. **Valve Engineer Jeremy Selan:** "The display behaves really well when it gets warm, it's actually a benefit. One of the things you want in VR is fast transition times, because we run at such high framerates—I haven't explicitly said it, but we work on the low end from 72 all the way up to 144Hz, and the best way we're able to get those fast transition times at higher framerates is through the warmth of the device. So you can go too far, of course, like a lot of heat—this is an LCD system, and a lot of heat in an LCD isn't good—but there's this happy medium in between where the device is warmer than ambient and can therefore achieve really fast times." Valve's justification for LCD over OLED came down to "FOV, eyebox, cost, persistence, [and] weight." To avoid blurring, the duty cycle for the LCD panels is a flat 8% at all framerates, meaning that at 120Hz with ~8.3ms frametimes, the illumination period is ~0.67ms per frame. Sound is provided via four 16mm drivers, two per ear, that have been tuned to cancel out vibrations. **Valve Engineer Joy Lyons:** "The thing that's actually really interesting is when you integrate audio into a sensitive piece of equipment like the core module where there's all sorts of tracking that's happening, the vibrations can be a limiting factor in terms of how loud the system can get. And so we designed it with two speakers per side, they're actually oriented in different [...] orientations, so that they can cancel the vibrations at the IMU in the core, while actually still doubling the amount of sound that the user is getting." As with thermals, we have a separate deep-dive coming up on acoustics. Joy from Valve had a lot of information to share here, so we’re including that with our separate thermal video. Headphones can be supported by Bluetooth and 3.5mm adapters. As for the controllers, Valve claims 40 hours of battery life, depending on the AA batteries used. * * The controllers look similar to the Quest 3's at a glance, although Valve emphasizes "input parity with traditional game pad," which is part of its push to play regular non-VR games in VR. Basically, you get all the normal controller buttons in the normal places. Also, the capacitive grip sensing has a limited gradient for the fingers and on/off sensing for the thumb. We haven't yet tested the Index for comparison. Valve taking a third stab at VR (and a second stab at SteamOS, and Steam Machines, and the Steam Controller, and arguably the Steam Link) feels justified in the wake of the Deck's success. As we understand it, Valve has been working on the Frame since at least the launch of the Index in 2019, according to one of the engineers we spoke to. That's interesting context, since it means Valve had the confidence to start work on another VR headset even before the pandemic VR boom, the launch of Half-Life Alyx. The Frame should be shipping to developers in the immediate future or already. ### **Steam Controller** The new Steam Controller, which is not called the Steam Controller 2, follows Valve’s previous Steam Controller that has seemingly been memory holed and overwritten for its name after being dumped en masse for $5. * * The controller uses Valve’s dual touchpad input and significantly repositions and reshapes the controller ergonomics versus the original Steam Controller. Valve noted that it benefits from the Deck’s popularity and showed that controller key mappings can transfer directly to the Controller. #### **Specs Summary** * * * * * The new Steam Controller can connect to anything running Steam or the Steam Link app and connects via the Controller Puck wireless, Bluetooth, or USB. The controller contains a lithium ion battery that can be recharged through the puck or USB, the thumbsticks are TMR and have capacitive sensing, there's a gyro with grip sensing, and what Valve calls "more advanced and powerful motors for high definition rumble" in the grips. The inputs completely replicate the inputs found on the Steam Deck, minus the touchscreen. #### **Details** * * The "Steam Controller Puck" is a dedicated USB dongle that connects to the device running Steam and provides dedicated wireless connections for up to four Steam Controllers per puck. This is both spiritually and literally a successor to the original Steam Controller dongle; Valve told us that the wireless protocol comes directly from the original Controller. Bluetooth is still an option, but it has downsides. **Valve Engineer Jeff Mucha:** “So if you add multiple devices they share the bandwidth, and then when that bandwidth runs out you have, you know, issues with lag and dropout.” **Valve Engineer Yazan Aldehayyat:** “But also, it's just that different PCs have different capabilities, different RF environments. It's more about the consistency than anything, like a device with Bluetooth will do really well, it's just that the consistency device to device and the consistency of the RF environment really comes into effect." Maybe it's just us, but one of the most interesting aspects of the controller is the antenna design. **Valve Engineer Jeff Mucha:** “One of the things for, like, ESD, when you have to go through compliance on a product, you have to zap it with the ESD gun, make sure that it's going to be good. With our antenna design, we landed on a design that solves the cost problem as well as ESD, because the antenna is basically grounded. So, if you hit it with ESD it doesn't blow out the front end of the chip. So, that was something that we learned on Index, and we brought that in for here. **Gamers Nexus:** “So, the cost and also just functional for the…” * * * **Valve Engineer Jeff Mucha:** “Cost, functionality, and it actually gives us, you know, we say we have a five meter range by getting the antenna away from everything else. The controller has a lot of stuff all packed in there, so where do you put the antenna where your hands aren't going to be? [...] So the antenna, you can see the thumbsticks are here, the antenna's there, so it's kind of right in the ideal spot where you're not putting your finger up here in normal gameplay." * * Valve claimed approximately 8ms of latency with a wireless connection through the Puck, but emphasized that performance shouldn't worsen as more controllers (up to four) are connected to it. The Steam Machine has the same functionality built-in, with a dedicated antenna (separate from the WiFi and Bluetooth antennas) for connecting up to four Steam Controllers. * * * The Puck also functions as a magnetic charger with pogo pins. Valve says its conservative estimate for battery life is 35 hours. The battery is a lithium ion pack. We asked about user repairability and replaceability of the battery. Valve noted that it's intended to be user serviceable in the sense that you can get to it with a screwdriver. In addition to the touchpad haptics already found in the Steam Deck, the Controller contains more advanced haptics. **Valve Engineer Yazan Aldehayyat:** "One example that helps us, think about like, speakers. So like, some speakers have really wideband performance that they'll play flat for a big frequency range. Some speakers don't, right, they're like a bass that only operates in kind of a very narrow frequency band. And so basically what we have here is something that's got… it's a linear motor, that has a wider bandwidth, so that it can replicate those higher frequencies [...] and waveforms better. So it's a better, higher fidelity speaker in some ways, just in a haptics world." Valve was clearly proud of its new Tunnel Magnetoresistance (TMR) thumbsticks, which are apparently the hot new successor to Hall effect sticks. **Valve’s Jeff Mucha:** "And you can see these thumbsticks, one of the things that, you know, Hall sensor thumbsticks have been a big thing in the industry; TMR is kind of new to the scene. It solves some of the power problems that Hall had, and keeping the sensitivity. **Gamers Nexus:** “Are these off the shelf? Are they customized for the sticks that you guys are using?” **Valve’s Jeff Mucha:** “We work with external suppliers, and we're not like fully integrating the thumbstick into, like the mechanical housings, but yeah we're using off the shelf suppliers for these." The Controller contains a 6-axis IMU, and capacitive grip sensing that can dynamically turn gyro controls on or off. The new Steam Controller’s LED markers allow it to be tracked by the Steam Frame when in VR. ### **Conclusion** Visit our Patreon page to contribute a few dollars toward this website's operation (or consider a direct donation or buying something from our GN Store!) Additionally, when you purchase through links to retailers on our site, we may earn a small affiliate commission. We have a lot more coverage to come, which includes a separate Steam Deck video. We will also have a video on the acoustics, power, and thermal engineering behind the Steam Frame. We’re also really excited to test all of this and are interested in seeing the OS hitting desktops outside of the Steam Machine. All of this hardware is on track for an early 2026 release date with pricing TBD. * * *

Our Most In-Depth Case Test Yet: HAVN BF 360 Flow Case Review, Fan Benchmarks, & Smoke Testjimmy_thang November 11, 2025 ## We analyze the HAVN BF 360 Flow’s thermal performance across a variety of scenarios, examine its fan flow performance, and bring smoke testing back The Highlights * The HAVN BF 360 follows-up the company's HS 420 that we previously reviewed and feels like a direct competitor to the Fractal Torrent * The company misspelled recommended as “recommendned” within the case * The HAVN BF 360 Flow is a high airflow, mesh-fronted case that charges a premium for its looks and build quality * Original MSRP: $160 - $190 * Release Date: October 2025 #### Table of Contents * AutoTOC Grab a GN Tear-Down Toolkit to support our AD-FREE reviews and IN-DEPTH testing while also getting a high-quality, **highly portable 10-piece toolkit** that was custom designed for use with video cards for repasting and water block installation. Includes a portable roll bag, hook hangers for pegboards, a storage compartment, and instructional GPU disassembly cards. ### Intro Today we're covering the BF 360, the second-ever case from HAVN, with two massive 40mm thick intake fans and a 30mm exhaust fan. This case feels like the first real follow-up to the Fractal Torrent’s (watch our review) impressive airflow. _Editor's note: This was originally published on October 25, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication._ * * * #### Credits * * * ##### Test Lead, Host, Writing Steve Burke ##### Testing, Writing Patrick Lathan ##### Camera, Video Editing Vitalii Makhnovets ##### Video Editing Tim Phetdara ##### Writing, Web Editing Jimmy Thang * * * * * * * * * * The BF 360 has big toilet-bowl airflow scoop, a plate that claims to prevent recirculation (which we tested), HAVN's overkill dedication to rubber accents on every single contact point, excellent cable paths, and it even has “recommendned” cable routing. That’s right. They recommendned it so much that the recommendnednation is repeated across the entire back of the chassis. HAVN can finally join other manufacturers for their typos, like Gunnir’s “Unkonw.” This is also the first case we've reviewed with a stone-style front panel, and we're ready for that change after two straight years of wood accents -- not that those have been bad, they’ve just been plentiful. The HAVN BF 360 is $160 without fans and $190 with the 3 custom fans included. ### **Overview** HAVN entered the market last year with the HS 420 (read our review), a brand-new concept from an experienced group of designers. The 420 had a lot of thought put into its airflow paths in particular, and its mechanical design, but it was still a glass-fronted case that shipped without fans. The HS 420 falls more into the pricier showcase category (alongside the HYTE Y70 (read our review) and Lian Li's O11 family) than the high-airflow category. Now, HAVN is back with the BF 360, a smaller and less expensive model with a fans-included Flow variant, directly targeting the high-airflow champions like the Fractal Torrent and Lian Li's Lancool cases. We still have plenty of praise for HAVN's attention to detail, but we've got a couple of notes to elaborate on later: first, HAVN didn't communicate a minor difference between the case fans and individually-purchased ones. Second, we identified a mistake on the active preorder page that said two fan hubs came with each case, which was wrong. It comes with one. After we noticed that, HAVN fixed it. The BF 360 is available in black or white, with the Flow listed at $190 and the non-Flow (without fans) at $160. Purchased separately, the Flow's two H 18 fans and single H 14 would total $70. #### **Specs (for Flow variant)** Form Factor| Mid Tower ---|--- Material| SGCC sheet metal, tempered glass, ABS, Nylon, Silicon, Zinc alloy, NdFeB magnets PC Case Dimensions (L x W x H)| 515 x 254 x 522.6 mm Net Weight| 14.79 kg Warranty| 1 year Notice: Warranty support is only available for orders shipped within the United States. Orders via forwarding services are not supported.Please review our full warranty terms for more information. Front I/O| USB 3.2 Gen 2 Type-C x1USB 3.2 Gen 1 Type-A x2Audio Jack x1Power & Reset Button x1 Motherboard Support| E-ATX (Max Width 277 mm) / ATX / M-ATX / Mini-ITX Expansion Slots| 7 Fan Support| Maximum Amount: 7 Top: 2 x 180 mm / 2 x 140 mm / 3 x 120 mmFront: 2 x 180 mmLeft: N/ARight: N/ABottom: N/ARear: 1 x 140 / 1 x 120 mm Fan PWM Repeater| Fan Ports: 1 x PWM repeaters, 6 ports totalLocation: 1 x Cable Management TrayFan Connectors: 6 ports per repeater x 4-pin PWM(Connect your primary fan to the white port to control PWM signal. All other ports mirror the white port's signal.)Fan Rated Voltage: 12VFan Safety Current: Up to 2.6A across 6 portsPower Connector: 1 x SATA PowerPower Safety Current: 1 x SATA connector power input, 12V → 2.6A (max) CPU Tower Cooler Clearance| Height: Up to 195 mm. Liquid Cooling Support| Top: 240 / 280 / 360 mmFront: N/ALeft: N/ARight: N/ABottom: N/A Graphics Card Clearance| Length: Up to 410 mmWidth: Up to 195 mm. GPU Support| 1 - 4.5 Slots RAM Clearance| If the radiator & fans thickness is greater than 80 mm: 280 mm - 40 mm / 360 mm - 60 mm Power Supply Support| ATXLength: Up to 200mm (Clearance - 260mm) Storage| Right Mount: 2 Storage Bays: 2 x 3.5'' / 8 x 2.5''Cable Management Tray: 1 Storage Bay: 1 x 3.5'' / 2 x 2.5''Storage Bay Included: 2x (2 x 3.5'' / 4 x 2.5'') Dust Filter| Front: 1 (Built-in)Top: 1 (Built-in)Rear Bottom: 1 (Removable) Pre-installed Fans| 2x H Series Fan - H 181x H Series Fan - H 14 _Specs copied from manufacturer materials, please read review for our own measurements and opinions_ ### **The Build** Let’s get into the build. * * Like the HS 420, the BF 360 has hard plastic cable channels that have been labeled with their "recommendned" uses and built-in velcro. HAVN probably wishes it had used stickers for those labels (like it did in the HS 420), but at least this will make TRYX feel better about the "discovered an unknow" typo. * * The predefined cable paths may come across as restrictive, but we have a hard time imagining a scenario where we'd want to route cables anywhere except the locations that HAVN chose. SATA power and data cables are the only ones that don't have a clear destination. * * * * * * We saw alternative front panel prototypes when visiting HAVN earlier this year, including a wooden one, but we're happy that HAVN steered away from the Fractal look and chose a less common stone-like pattern. They’ve beaten Cooler Master to the punch on this as well. The speckled plastic finish reminds us of an old keyboard. On that subject, all of the case fans are black in the black model, while in the white model the front fans are black and the rear fan is white. We think this is an intentional choice to hide the front fans, but it's something to be aware of if you want to color-coordinate your build. * * The whole front panel assembly hooks into the case toollessly and is held in place with magnets, similar to the Fractal Torrent. The BF 360's front filter layer can't be separated for deep-cleaning as easily as it can be in the Torrent, but the stone-textured layer can be removed with six screws. * * Inside the case, there’s a scoop. We've been calling these airflow scoops "toilet bowl" designs even though the NZXT H5 Flow is the only one that really looks like that. We've also seen examples from Fractal and Corsair, but unlike each of those, the "air guide ramp" at the front of the BF 360's PSU shroud is made of metal and is permanently attached with rivets. It’s also relatively flat, which HAVN’s thermal engineer says tested the best in their A/B testing. Usually scoops like this are removable for radiator compatibility reasons. The Flow doesn't claim any front radiator compatibility in the first place. When planning a build in the non-Flow, keep in mind that the front edge of the scoop is 8cm back from the front fan bracket, although that's plenty of room for conventional radiator and fan combos. * * To match the scoop, there's an "Anti-Backflow Plate" that blanks out the part of the top fan mount above the 180mm front fans. "Backflow" in this case would mean air flowing out of the case through the gap and recirculating into the front fans, reducing their cooling effectiveness. * * * * The GPU support arm in the BF 360 doesn't match the quality of the case's design. The HS 420's support is sturdy and braces against the PSU shroud, while the BF 360's is less reinforced and relies solely on friction with the rubber cable grommet to keep from sliding down. Moving to two points of contact with the GPU is a good idea for an upgrade, but in practice it means that one point will always be looser than the other. The most positive change is that the new design can provide support to about ~35cm down the length of a GPU, whereas the HS 420's support can only reach ~28cm. * * * For packaging, HAVN arranged the styrofoam inserts so that they don't catch on the punched-in handles of the outer box, which is surprisingly rare. It’s not a big deal, but when you unpack multiple cases every month for years on end, you start to notice these things. HAVN also uses labeled ziploc bags for its screws, which helps reduce waste (versus single-use bags) while keeping screw types separate. On one of the two review samples we were sent, the front I/O wasn't lined up correctly with the cutouts on the surface of the case, making the USB Type-C port inaccessible. This is the first fit-and-finish problem we've seen from HAVN. Fortunately, we were able to fix it ourselves, but this shouldn’t happen on a nearly $200 case. * * * * * * * HAVN's H Series fans have a few features worth calling out. First, the unusual shape of the hub was supposedly intentionally designed to "guide air towards the fan blade for smoother flow pattern," while the fan blades themselves have a sawtooth pattern that we've seen many times in the past from companies like Fractal and EK. The fans are 40mm thick, which gives extra length to the fluid dynamic bearings, which HAVN claims increases durability and stability. The fan frames are molded so that cables can be threaded through the corners, which is a nice way to reduce cable clutter without committing to something like Lian Li's UNI FAN system, and the rear fan's cable is routed through a special hidden channel at the top of the case. Retail H-series fans include the "Loopin™ Cable System" with daisy-chainable male and female connectors on each fan, but after we identified a potential error in the documentation, HAVN clarified that the ones shipped with the BF 360 Flow all have regular old 4-pin connectors. * * * * * We're not sure how to feel about the trays for storage drives. They look like a nightmare to manufacture, with rubber grommets, rivets, studs, and a removable inner piece, but the end result is a completely normal drive capacity: four total 2.5" drives, or two total 3.5" drives. The two trays are mounted in the cable management area by default, but there's actually a third mounting location behind the motherboard tray that may be preferable if you want to free up room for cables. * * * HAVN has remained committed to putting rubber damping on every single point where two movable metal surfaces would contact, and this extends to the drive bays, which is the single area in the HS 420 where we noted a lack of rubber. There are also textured grips inside the handles on the side panels. They don't do much functionally, but from a UX perspective they indicate where to pull on the panels and in what direction. * * * * As for radiators, the top radiator tray is the entire roof of the case, so removing two screws and sliding the tray off gives unrestricted access to the interior. The only downside we noticed is that if you mount HAVN's extra-thick 180mm fans to this tray, they'll obstruct the cutouts along the top edge of the motherboard, and CPU power cables may come close to the fan blades. HAVN's claims for radiator support are extremely conservative, with the Flow only officially supporting radiators in the top of the case, and only up to 360mm. * * * * * The front and top panels are held on with magnets, but all panels have additional screws to secure them during shipping. The glass side panel is beveled on each corner, and the plastic gutter that it seats in is shaped to match. Both side panels have metal studs on the top edge (in addition to the usual snaps) to make sure nothing sags out of alignment, which is similar to the strategy TRYX used for its L70 (read our review) update. ### **HAVN BF 360 Flow Thermals** Grab a GN15 Large Anti-Static Modmat to celebrate our 15th Anniversary and for a high-quality PC building work surface. The Modmat features useful PC building diagrams and is anti-static conductive. Purchases directly fund our work! (or consider a direct donation or a Patreon contribution!) Other than the normal suite, we've selected three alternate configurations for testing: one with the anti-backflow plate removed, one with two additional H 18 (180mm) fans added as top exhaust, and finally another test with the additional 180mm fans but with one as intake and one as exhaust. According to HAVN (and our own experience), top intake in front of the CPU cooler and top exhaust behind it should help to avoid stealing air from the cooler. The scoop in the BF 360 isn't removable, so we tested it as-is, but we've done experiments in the past with removing or taping off airflow scoops at the front of PSU shrouds. If you're interested, you can see the most recent results in our Meshify 3 review. HAVN's own HS 420 is an obvious comparison to make, but they don't sell a version of that case with fans included (yet), so our thermal results come with a caveat that they aren't truly "stock." HYTE competes closely with HAVN as a company, but we can't make thermal comparisons against the recently-launched X50 ($160 without fans) until our review is finalized. We'll judge against HYTE's $180 Y60 instead, but the Y60 isn't a high airflow case. * * * As for comparisons that we feel good about, Fractal's Torrent and the much newer Meshify 3 are on the list. Lian Li's Lancool 217 is there as well. The top-tier Meshify 3 Ambience Pro RGB has dropped to a competitive $180 on Amazon, while Lian Li continues to offer brutally low prices with the 217 at $122-$125. #### **CPU Full Load Thermals - Full Speed** Before we get to noise-normalized testing, let’s start with full speed testing to better evaluate alternate case configurations at full speed. Noise-normalized testing is better for comparing different case models, but running the fans at full speed makes it easier to compare configurations of a single case. With the fans at full blast, the BF 360 Flow averaged 37 degrees above ambient for the all-core CPU temperature and 41 degrees for the P-cores alone while at 44.5 dBA from 1 meter in our chamber. Removing the anti-backflow plate caused the CPU temperature to rise by just over one degree, so it seems like HAVN was right to include it. It works. Adding two 180mm exhaust fans to the top of the case made CPU temperatures worse, which is an expected result that we have continually re-proven for about 12 years now. This is because the top-front exhaust fan steals air from the CPU cooler before it ever hits it, so it’s a relatively simple concept. Overall, the change wasn't significant. On the other hand, flipping the frontmost 180mm fan to intake also reinforced findings from our previous test bench iteration, resulting in the lowest averages we've recorded in this test. The results dropped down to 35 degrees all-core and 38 on the P-cores. These are excellent results, although this was with adding two extra 180mm fans, so adding similar fans to other cases would likely give the BF 360 some company. If you want to go crazy on airflow, doing this 4x 180 configuration might be a fun way to do it. The stock BF 360 Flow tied the Lancool 217 (read our review) (at the same dBA) for CPU thermals, making it the best on this chart other than the Lancool 216 (read our review) (at 42 dBA). Peak performance of the thick H 18 intake fans is high, allowing the BF 360 to surpass the Torrent and Meshify 3's 39 degree all-core average. Like the Torrent and 217, though, the BF 360 is noisy at max speed, registering 44.5 dBA. #### **GPU Full Load Thermals - Full Speed** The stock BF 360 Flow averaged 40 degrees for GPU thermals and 44 on the VRAM, which was completely unaffected by the backflow plate. That makes sense since the plate is at the top of the case. Adding the two 180mm fans as pure exhaust to the top of the case aligned with the GPU's flow-through cooler and brought down the average to 39 degrees, while the mixed intake-exhaust configuration raised the GPU's average to 42 degrees. This is why it’s important to consider more than one temperature during configuration: This was better for CPU thermals, but is worse for the more sensitive GPU where you can get slight boosting benefits from a few degrees. With a flow-through GPU cooler pushing hot air up towards the roof of the case, the best compromise may be to leave the stock configuration alone and stick to front intake, rear exhaust. The 217 is tied with the BF 360 here, just like it was for CPU thermals, and the Meshify 3's relative placement is also the same at one degree warmer than the BF 360. Meanwhile, the Torrent is slightly cooler at 39 degrees. #### **CPU Full Load Thermals - Noise Normalized** Noise-normalized thermals are next. For these tests, we place the case in our hemi-anechoic noise chamber and lower case fan speeds in tandem until the overall noise level hits our desired 27 dBA SPL threshold. The BF 360's large intake fans and unique fan hubs should help it in noise-normalized testing. The BF 360 Flow is in excellent company here, with an average all-core CPU temperature at 41 degrees Celsius above ambient and P-cores at 45 above ambient. That matches the Lancool 217's performance almost exactly (whether the 217 is in CPU or GPU mode), and although the Meshify 3 Ambience Pro RGB was cooler, it's within our rough one degree margin for error. The Torrent is similarly close in the other direction, with the BF 360 coming out ahead. The HS 420 result shown here is tied with the Torrent, but that's with the 420 using the Antec Flux Pro's stock fans since it didn’t ship with any. That’s 6 fans in the HS 420 for that test. The Y60's stock performance is on the complete opposite end of the chart at 53 degrees all-core, but again, the Y60 isn't an airflow-focused case. The Flux Pro (read our review) ends up about the same as the BF 360 as well, so overall, the BF 360 is one of the most competitive cases on the chart and is more or less within re-run variance of the best cases. It is functionally tied for the best. #### **GPU Full Load Thermals - Noise Normalized** The BF 360's noise-normalized GPU temperatures in the same test were also very good, at 43 degrees above ambient and 47 for the VRAM. The Torrent keeps the crown here with its 40 degree average, which is a meaningful improvement over the BF360. The BF 360 falls within one degree of the Lancool 217's GPU mode result, and it almost exactly matches the Meshify 3's average. As with CPU thermals, the HS 420 is one degree warmer than the BF 360 (using Antec's fans). That’s a good result for the BF 360 considering the configuration we had the HS420 in. The stock fans are biased low so that the lower intake fan points into the airflow scoop, which shows an intentional focus on GPU thermals. Overall, the BF 360 keeps pace with its competitors, but it's worth keeping in mind that the 217 in particular is significantly cheaper. HAVN’s BF 360 does well overall here and is in the top set of results, but didn’t top the charts like we’d hoped based on its fan configuration, but it’s at least in good company. #### **VRM & RAM Full Load Thermals - Full Speed** The configuration with one top intake and one top exhaust was the best for VRM thermals, with a chart-topping average of 18 degrees above ambient. It was also by far the best for DRAM thermals, with a single-digit 9 degree average. It makes sense: the intake fan pushing cool air directly into the memory and the exhaust fan pulling hot air directly out of the VRM heatsinks is the ideal scenario for those particular hotspots. ### **HAVN BF 360 Flow Smoke Flow Testing** Smoke testing is challenging and poses problems. We’ve been experimenting with this for over a decade now, and the first time we tried it, we used plumbing smoke pellets. It worked great for visualizing airflow, but it filled the area with carcinogenic smoke. For this round, we’re using a fog machine instead. The fog is safer, but it comes with a tradeoff: residue. Fog oil coats everything it touches, leaving a thin, greasy film on all the components. The system will still run fine electrically, but it’s gross and isn’t something you want on your main test hardware. Because of that, we’re using a different PC for smoke testing instead of one of our benchmark systems. The next problem is how you get the fog into the case. The common approach is to point a nozzle at the front intake and spray smoke in, but that’s a bad idea. As soon as you add velocity and essentially turn the fog into a mini air jet, you’re changing how the case actually breathes. Wherever you aim that nozzle, you’re forcing the smoke in unnaturally, which tampers with results. Ideally, you’d have the PC surrounded by an even cloud of fog so that it can pull in air naturally. That’s the only way to see what’s really happening with positive or negative pressure inside the chassis. About ten years ago, we built a homemade smoke chamber out of cardboard, tape, and Saran Wrap. It actually worked well, but it took five hours to set up every time. * * * * * This new setup is a streamlined version of that old idea. We’ve built a semi-sealed box, which we put in front of the PC. It’s not perfectly airtight, but it’s close enough. From there, we fill it with fog from the top. The case is then left to pull in air on its own when we remove the panel, with minimal interference from us. That way, we’re not skewing the results by blasting smoke at specific points. As for the fog itself, it’s not ideal, but it’s the best middle ground. The oil residue is annoying, but it’s far better than from plumbing pellets. We’re using parts we don’t mind getting a little messy and it’s still representative of our test system. It’s an early iteration of the setup, but we’re already planning upgrades, like better seals and gaskets for the fog hose to make it more consistent. This approach gives us a realistic, repeatable way to visualize case airflow without interfering with it. * * The front-top fan appears to spray wide, with the top of the inlet arcing air up toward the top of the case. * * * We can see air from the top half of that bottom fan entering the case and immediately going through the flow-through fan. Beneath it, the rest of the air hits the scoop, and that air makes it through to the left-most GPU fan, which will enable cooling into the GPU. Some air exits straight out of the case via the PCIe slots, so we're not getting the negative pressure effect where air is fed in through the back, as we saw with the old NZXT case. * * Meanwhile, the air going through the flow-through area is deflecting the top intake air up and away from the CPU cooler. This matches our testing during the 30-series of NVIDIA’s flow-through coolers. The CPU cooler ends up a little warmer as a result, and so does the CPU temperature, but this is the nature of flow-through coolers and not of the case itself. CPUs are low enough power compared to GPUs that they can typically handle it without meaningful change. As a result of the GPU's giant flow-through area, the case intake is being deflected up and toward the roof of the case. This is why GPU temperature goes up with a top-front intake fan, but CPU temperature goes down. * * We also noticed a slight recirculation pattern in the top front. This isn't a problem since it's eventually finding its way out, but this is a good further defense for why the back-flow plate is present. This air would be exiting the case and re-entering the fan without that plate, and it'd have been warmed in the process. It's not perfect, but people have wanted us to bring these back for years now, so we did our best to resolve our hangups with it with this quick mock-up. We have more we can add later. ### **Fan Performance Testing** * * * * We also ran some quick fan tests on the fan machine. We’re slowly starting to roll this data out. As we talked about in February this year, we’re just going to have to accept that we need to begin building a data set and it won’t be everything we want all at once. This will also be redone and improved upon over time. We want to caution you that we are inexperienced in both presenting and collecting this data, and so it does not have our same high confidence as other data we collect. Although we think the data is accurate, there is a higher chance of an inaccuracy. These charts will be marked as experimental for that reason. We’re just comparing two fans here today and without their cases. The point of this test right now is to better put data to the thermal results of the BF 360 vs. the Fractal Torrent, since they’re close competitors on the charts. The first test shows the maximum CFM on the machine. At full speed and max CFM, the BF 360 fan reached 169 CFM. This was at 1279 RPM. The Fractal Torrent Dynamic X2 fan hit 141 CFM with its 1187 RPM, so slightly slower, but a somewhat large gap in flow performance. The Torrent fan is older and a couple millimeters thinner. Looking back at the full speed case chart, with all factors including case design considered, the BF 360’s 41-degree P-core average had it ahead of the Torrent by about 2 degrees. A lot of that is attributable to the fan here rather than just the case design. Looking back to the fan chart: We also took case noise-normalized numbers. This means that we ran the fans at their speed we ran them in our noise-normalized case test; in other words, the fan is not noise-matched standalone, because we didn’t test it that way thermally. Matching the case conditions gives us better insight as to differences between the case results. When running at noise-normalized case levels, we measured 85 CFM on the HAVN BF 360 and 72 CFM on the Torrent. The BF 360 fan is testing better in our initial tests here. The noise-normalized CPU numbers for the cases had the BF 360 at 45 degrees to the Torrent’s 46. We had some pressure numbers as well, which we’re only presenting at full speed. This is just because we’re still studying how we want to present the rest of it. But for the full speed results, the Fractal Torrent had a 2.31 mm H2O for pressure performance. And the Haven BF360 was at 2.87. That is for the maximum mm H2O at full speed, which would mean the minimum flow in that test. Anyway, we have a lot more to do here, but the simple matter is that we can’t figure out what that is without just starting to build a data set, so that’s what we’re doing. This is a very rudimentary start for you all. It’s easy to forget our other testing has years of experience and collection behind it, so we’ll get there eventually with this. ### **HAVN BF 360 Flow Conclusion** Visit our Patreon page to contribute a few dollars toward this website's operation (or consider a direct donation or buying something from our GN Store!) Additionally, when you purchase through links to retailers on our site, we may earn a small affiliate commission. The $180-$200 range where the BF 360 Flow lands has become the most crowded segment over the past several months of our case reviews: currently the Antec Flux Pro is $180, the Corsair 5000D RS ARGB is $200, the Meshify 3 Ambience Pro RGB is $180 depending on sales and availability, and the MasterFrame 500 Mesh we just reviewed is… $150, despite launching at $190. Of those, the Meshify 3 in particular falls into the same category as the BF 360, in the sense that it's a high airflow mesh-fronted case that charges a premium for looks and build quality. Fractal has the advantage of cheaper Meshify 3 case SKUs that still come with fans, although we found the ring of LED lighting in the Ambience Pro RGB to be a big part of that case's appeal. Fractal also still sells the Torrent, which bears clear similarities to the BF 360 with its two extra-large 180mm intake fans and excellent thermal performance. Cable management is better in the BF 360 than the Torrent, but you can't go wrong choosing between the two, although it does sting to pay more than the original $190 MSRP for the Torrent when the case is more than four years old. Alongside Fractal, HYTE is the major competitor for HAVN, but the HYTE cases that we've reviewed don't make sense as alternatives to the BF 360 Flow. You may want to wait for our upcoming HYTE X50/X50 Air review, since those cases are more focused on thermal performance in a way that the HYTE Y-series (currently) isn't. Neither of the X50 variants come with stock fans, but the X50 non-Air is priced at $160, which matches the fanless BF 360. In terms of raw performance per dollar, Lian Li still has things locked down. There are other factors that matter when purchasing a case, but if all you care about is minmaxing, the Lancool cases are nearly impossible to beat. The older 216 was more similar to the BF 360 Flow with its two wide-framed 160mm intake fans, but this year's Lancool 217 performs just as well, and it's still readily available at its launch MSRP of $120. Although it's not quite in the same category, we also have to mention that the Lancool 207 is priced insanely low at $83-$90 while easily keeping up with the BF 360's thermal performance. Given the HS 420's price increase to $230 (minimum) and the fact that it doesn't come with fans, HAVN has successfully avoided competing with itself. The BF 360 is exactly what we wanted after reviewing the HS 420. The 420 is big, it's got a lot of glass, it doesn't come with any fans, and it's over $200 (although the price makes sense for what it is). The release of the HS 420 left an opening for a smaller, less expensive, more airflow-focused case, which is the BF 360. The build quality is high, the thermal performance is extremely good, and although the Lancool 217 and other Lian Li models may have more favorable dollars-to-performance ratio in a strict sense, the BF 360's visual appeal and overall refinement make up the difference. Other than the “recommendned” cable paths, of course. * * *

Intel's New GPU: Xe3 Architecture Changes, Handheld Gaming CPUs, & XeSS3jimmy_thang October 29, 2025 ## Intel announced its Panther Lake mobile solution with Xe3 GPU architecture updates, building upon Xe2 and moving toward an eventual Celestial launch in the desktop GPU space The Highlights * Xe3 is distinct from Celestial * We suspect that Xe3 sets the foundation for some of what Celestial will eventually get * Intel's Panther Lake updates also include claimed power management and balancing improvements for CPU vs GPU power budget allocation in gaming scenarios #### Table of Contents * AutoTOC Grab a GN Tear-Down Toolkit to support our AD-FREE reviews and IN-DEPTH testing while also getting a high-quality, **highly portable 10-piece toolkit** that was custom designed for use with video cards for repasting and water block installation. Includes a portable roll bag, hook hangers for pegboards, a storage compartment, and instructional GPU disassembly cards. ### Intro Intel today is announcing its Xe3 GPU architecture to follow-up its Xe2 architecture. It’s not Celestial yet, but it’s getting close. _Editor's note: This was originally published on October 9, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication._ * * * #### Credits * * * ##### Test Lead, Host, Writing Steve Burke ##### Video Editing Mike Gaglione ##### Writing Jeremy Clayton ##### Writing, Web Editing Jimmy Thang * * * * * * * The biggest change to Xe3 is that it’s just larger, with render slices scaling up to more Xe cores per slice, an increase in L1 cache from 192KB to 256KB, a significant increase in L2 cache, and more registers that are better utilized. Micro benchmarks show significant improvements in occluded primitives culling for unnecessary triangles when rendering game scenes in addition to improvements in anisotropic filtering. Its variable register allocation and register changes also aim to unclog the pipeline so that the hardware can be better utilized, as one of the biggest problems with Arc in its current Xe2 and Battlemage implementation has been that there’s plenty of hardware, but it’s not getting used properly. This is a mix of both hardware issues, like with fixed function units in the architecture, and driver issues, which it has been slowly addressing. Some of this included moving off of emulation of things like execute indirect previously to eliminate overhead. For Xe3, Intel noted to us some of its driver improvements and software control panel focus as well, all of which should benefit the company as it moves toward its eventual dGPU Celestial GPUs. This accompanies a number of other announcements related to its Panther Lake mobile solutions and laptop hardware, plus some “AI” and NPU hardware. We’re mostly going to focus on the IP block of Xe3 and the architecture and won’t be as focused on the product side for laptops. Although this isn’t a dGPU part, it’s likely that this approach will either be directly found in the next dGPU or will at least indicate which direction Intel is going. Intel was clear that this isn’t exactly Celestial, which is the architecture following in the Alchemist - Battlemage - Celestial - Druid lineup. Intel noted that “Xe3P” will follow Xe3. The “P” unironically stands for “Plus,” showing old Intel habits die hard. Intel didn’t confirm this, but the impression we got is that Xe3P will be the “real” Celestial GPUs, while this Xe3 makes major changes that likely set the stage for it. ### **Overview of Announcements** Intel had a lot of announcements to share with the press for today. For our coverage, we’re focusing almost entirely on the Xe3 changes and micro benchmarks. We’ll cover some of the other news as well, like performance/Watt improvements and XeSS changes, but we’re not going to get into the NPU and AI processing changes today. There’s enough to talk about just with the stuff that’ll affect consumer desktop components in the future (plus the immediate impact to laptops). All of this follows the announcement that NVIDIA is investing in Intel to build its own mobile parts with them later, but there’s no news on that topic today. This is all Intel’s hardware. #### **Naming Confusion** Briefly on the naming: Intel admitted its naming mix of Xe for IP and Alchemist / Battlemage / Celestial / Druid for branding has been confusing. It was careful to note that these parts are not Celestial and the impression we got was that they don’t want to burn the name on an incremental improvement prior to a pending major overhaul. Intel is sticking with “Arc B-Series” for the Panther Lake mobile parts, but is moving to the Xe3 architecture. Xe3P will likely be Celestial or desktop parts later. #### **Xe3 IP GPU Block** Intel specifically mentioned designing Xe3 to scale to larger configuration sizes, which would be good news for anyone who wants to see something higher-end than a B580-class card in the future. Let’s get into micro benchmarks first, then look at the block diagram. This is a chart of micro benchmarks, which are workloads designed to target extremely specific functions or behaviors on a product. A 2x improvement here won’t equal a 2x improvement in most real-world applications, but these allow us to see where the improvements are appearing. Intel published these for Xe2 also. In Xe3 for “depth writes,” Intel says it saw a 7.4x relative performance improvement normalized to clock frequency. We’re not certain, but our understanding is that this is not isolated for configuration size. This means that this isn’t a perfect comparison since the Xe core count is different between Xe2 and Xe3 in these tests. This 7.4x improvement outstrips the change in configuration size, though. We asked Intel what “depth writes” means. The company told us that it’s related to high-Z culling and that this bar represents better primitives culling in the pipeline, meaning culling of unseen triangles and geometry sooner in the pipeline so as not to waste resources rendering unseen objects in-game. An example might be if a building is obstructing a player -- there’s no point rendering the player if it can’t be seen. Culling isn’t new and batching primitives in ways that eliminate occluded primitives has been around forever, but this shows that there’s still plenty of ground to gain here for Intel. This will result in better utilization of resources and allocating them to more productive work. Intel told us that the improvement to this process is disproportionately beneficial, meaning that it should have an impact in gaming performance that would be more noticeable than other improvements. We’d expect this to carry over to future Celestial dGPU parts as well. The “High Register Pressure Shader” section also saw a large uplift in micro benchmarks at 1.9x to 3.1x. Scattered reads improved by 2.7x on the relative scale of time, with Intel noting to us that this has to do with using samplers to read data scattered across something like a texture (as opposed to a well-organized data set). Mesh rendering is also shown here, with Intel telling us that Xe2 had already provided a proof of concept around improving mesh shading. Intel noted that this micro benchmark is representative of workloads where a lot of polygons are present, telling us that the uplift comes from a larger cache and more efficient use of its registers. Culling also contributes. Quickly, Intel also saw uplift in anisotropic filtering, which is the old function that helps improve smoothness of textures and objects proportionate to the view frustum’s angle. Ray-Triangle intersection also improved by 2x in the microbenchmarks on the relative scale, which is noteworthy since Xe2 already benefitted from relatively large ray tracing improvements. Looking back at the Xe2 micro benchmarks, Intel then highlighted Draw XI and Compute Dispatch XI primarily. At the time it talked to us about this chart, Intel told us that this was due to implementation of native execute indirect support for indirect draw and dispatch, as opposed to its Xe1 emulation of these functions. ##### **Block Diagram** Time to get into block diagrams for how the new Panther Lake Xe3 block is constructed. This shows a 12 Xe-core configuration as the maximum size announced for mobile, with this configuration carrying 16MB of L2 cache, 2 geometry pipelines, 12 samplers, and 4 pixel backends. The L2 cache is noteworthy here. This is the new Xe3 render slice. A render slice is Intel’s terminology that defines a block on the GPU containing Xe cores. For reference, the Battlemage B580 with Xe2 has 20 Xe cores on 5 render slices, so each slice is just one part of the total GPU. The Xe2 slice had 4 Xe cores each, with Xe3 moving to 6 Xe cores per render slice. Intel also intends to scale-up the configuration size on mobile devices to a maximum of 12 Xe cores (or 2x render slices, up from 8 Xe cores on a prior 2-slice configuration). The Xe3 render slice shows that each Xe core has 8 vector engines, which is unchanged from Xe2 cores; however, Intel is increasing the cache size in Xe3. Intel’s Tom Petersen stated, “The first thing we’ve done is increase the size of our L2. By increasing the size of the L2 from 8MB to 16MB, we reduced the traffic that hits the memory interface. That’s important because the memory interface is typically one of the most precious resources on a graphics chip. We can see anywhere between 17% and 36% traffic reduction heading towards memory, which has a significant performance effect on these different applications.” Looking at Intel’s first-party results, it presents the improvement in the form of relative traffic on the SoC fabric (in the vertical axis) against a baseline 8MB L2 cache. Cyberpunk with RT showed a 19% reduction, Black Myth rasterized showed a 36% reduction, and the rasterized Steel Nomad test showed a 17% reduction. Intel also told us that it has increased its L1 Cache by 33%, noting a move from 192KB to 256KB. When we asked Tom Petersen which area of uplift he thought had the most impact on overall performance, he pointed us toward the register and thread changes. Intel has increased thread count upwards of 25% depending on configuration and has moved to variable register allocation. Petersen noted that occupancy of the compute units (including on Battlemage) previously wasn’t always high, despite them being available for work, meaning that there was more GPU hardware present than was being properly utilized by applications. Intel has focused on this in both drivers and hardware. He noted that previous register allocation and thread count choices would “starve the pipeline if the shader used too many registers,” which is being addressed. The ray tracing unit also got improvements. Intel says it “slowed down dispatches of new rays while the sorting unit catches up,” citing out-of-order dispatch and triangle testing. The ray tracing unit improvements seem to be largely attributed to asynchronous dispatch-test processes. Intel also highlighted a new URB manager as part of its fixed function enhancements, which is also where we find the anisotropic filtering uplift. Petersen stated this, “We also now have a new URB manager, which allows partial updates versus flushing the whole thing. Our URB is a structure where we pass results between our units inside of our GPU. It used to be somewhat of a serializing point; now we can actually use that partially without flushing each complex.” ### **Frame Inspection** Grab a GN15 Large Anti-Static Modmat to celebrate our 15th Anniversary and for a high-quality PC building work surface. The Modmat features useful PC building diagrams and is anti-static conductive. Purchases directly fund our work! (or consider a direct donation or a Patreon contribution!) We thought these next couple slides were pretty interesting as well: Intel showed a frame on Xe3 versus Xe2. These are not normalized for configuration size, so it’s not a perfect comparison and it shows a 12-core vs. 8-core configuration, disallowing a perfect like-for-like inspection. This is iso frequency and power, so it is at least normalized there. The horizontal axis is for API call execution, with the vertical axis being milliseconds of time to execute across a single frame being drawn (higher is worse). This is for Cyberpunk 2077. Of note, Intel shows an 8ms reduction to Xe3 with the compute and pixel shader section toward the end, assigning some of that uplift to the change to the variable registers and L1 cache size increase. We can also see that, according to Intel, the L2 benefits the render base pass with a 0.39ms improvement, preceded by the move to 10 threads (and variable registers) providing a 2.93ms improvement in the pre-pass. More broadly, Petersen told us in a call that the register allocation and number of threads would starve the pipeline if the shader used too many registers previously, which is being partially addressed here. He said that the previous architecture could cause a reduction in the utilization of available compute resources due to regular flushing of the pipeline due to regular reallocation into memory. This image is pretty cool and is a look at what actually happens in a frame when it’s being drawn. We have a full video talking about this previously. ### **Power Delivery** Intel’s focus on power delivery and power management cites learnings from the MSI Claw (read our review) devices and mostly comes in the form of ensuring proper resource allocation for power budget between the CPU and GPU, which should benefit laptop and handheld devices that have a limited power budget split between the two. Intel noted that previously, a lack of application awareness meant that the device could sometimes divert too much power to the CPU, leaving the GPU bottlenecked on its power limit while the CPU offered a level of performance that wasn’t being kept-up with by the GPU. Intel gave the MSI Claw as an example of a time this didn’t go well. The company noted that it improved on this earlier in the year with its Intelligent Bias Control v2 and is now introducing a v3 to build upon that. Because the system was previously unaware of the application being run, in this case a game, Intel said that software and hardware wouldn’t correctly balance the workload between the CPU and GPU, resulting in stuttering due to being power starved. “Intelligent Bias Control v2” took GPU heuristics and utilization metrics to then inform thread scheduling and resource assignment at the operating system-level. Intel had previously marketed improvements to 1% and 0.1% low metrics via better frame interval pacing as a result of this change. The new v3 version of this adds E-core first scheduling, which is self-explanatory in that E-cores get scheduling first when gaming. This sounds worse, and typically would be, but Intel says that the end result is reduced power diversion to the CPU by using lower power cores prior to P-cores, freeing-up more of the shared total power budget to go toward the GPU instead. In GPU-bound scenarios, like many games particularly on handheld devices, this is a better outcome than burning power on a component that isn’t as burdened. This comparison between Panther Lake and the prior generation of this bias control solution shows that peaks in power utilization have smoothed-out while the GPU power consumption has leveled to be more predictable. Reminder: This is a first-party tests. The GPU is also getting more total power budget as a percentage than previously, while reducing CPU power in exchange. For GPU-bound scenarios in particular, this should be a better outcome. It might help in some CPU-bound scenarios as well. ### **XeSS Multi-Frame Generation and Other Changes** * * Intel also announced XeSS 3, which includes XeSS-Multi-Frame Generation (or XeSS-MFG). A few more letters and they’ll have the whole alphabet. XeSS-MFG is conceptually similar to NVIDIA’s MFG. XeSS-MFG takes 2 real frames to calculate optical flow networks using motion vectors and the depth buffer, then uses that information to generate up to 3 frames between the 2 real frames. The frames are then displayed in order and paced in a way to minimize animation error. We also have a separate deep-dive on our new animation error testing methodology. The new “XeSS Frame Generation Override” setting in the driver software allows the user to set 2x, 3x, or 4x mode. Intel presented a few timelines of a single frame. One at native, and then several with various levels of XeSS technology in use. The shorter the duration that the frame is on the X-axis, the less time the frame took to complete. The top half of each shows instructions and the bottom half shows when the geometry pipeline is active. Compared to native, the raster, RT, and denoise sections of the frame are shorter on the XeSS 3 timeline due to rendering at a lower resolution. The first purple section represents XeSS-SR to perform the upscaling. The second purple block starts with the optical flow portion of frame gen, followed by 3 frame generation operations. Visit our Patreon page to contribute a few dollars toward this website's operation (or consider a direct donation or buying something from our GN Store!) Additionally, when you purchase through links to retailers on our site, we may earn a small affiliate commission. It seems like Intel’s argument is that the entire frame gen process takes less time than drawing one real frame, and is therefore better or something, but this totally ignores image quality. We’ve shown with both AMD FMF and NVIDIA MFG that the image quality sacrifice isn’t always worth it. Sometimes it is, but it’s not always as simple as being that way. Intel stated that these frames upscaled with XeSS-SR are the same quality as native, which is unlikely. Intel stated: “That frame is as good as the prior picture, the native frame. But it’s actually being run quicker.” We doubt this will be broadly true and will evaluate later on dGPUs. It was bullshit when NVIDIA claimed it, too. The quality can be good, but is rarely as good. Intel had some other side-by-sides that we take issue with, and that in combination with still having watermarks on the video means we’ll skip them and just test it ourselves later. Intel referred to the frame gen process as looking into the future. NVIDIA CEO Jensen Huang has said similar things about NVIDIA’s frame generation. Both of them are wrong, because all current methods of frame generation rely entirely on finished frames and engine data. These frames already existed and could have been displayed instead of holding them to run the frame generation in between. That isn’t looking into the future, that’s interpolating between two sequential snapshots of the present or near present. Until a predictive method of frame generation comes out, none of these technologies look into or generate “the future,” they at best interpolate the past. And that’s fine, but we’d really like it if these companies could get their shit together and stop saying that they generate the future. MFG represented on benchmark charts has been a major and ongoing controversy and misrepresentation of performance on NVIDIA’s side of things. Intel committed to relying on base raster performance without frame generation as the baseline for performance and said that, when it publishes numbers including upscaling or frame gen, those will be provided as supplemental to the base metric. We think this is a better balance of promoting the capability without totally misrepresenting the reality. Intel also talked about a new version of PresentMon that includes a few changes, partly accounting for frame generation technology. * * *

Excellent Build Quality: Cooler Master MF500 Case Review & Modularityjimmy_thang October 22, 2025 ## Cooler Master’s MF500 case has overall high build quality with good materials and sturdy panel, although it does have its downsides and areas we're critical of The Highlights * The MF in Cooler Master MF500 stands for Master Frame, which continues Cooler Master's redundant naming * The case has middling thermal performance * Cooler Master is back on its A-game with the MF500 series of cases * Original MSRP: $150-$190 * Release Date: October 2025 #### Table of Contents * AutoTOC Grab a GN Tear-Down Toolkit to support our AD-FREE reviews and IN-DEPTH testing while also getting a high-quality, **highly portable 10-piece toolkit** that was custom designed for use with video cards for repasting and water block installation. Includes a portable roll bag, hook hangers for pegboards, a storage compartment, and instructional GPU disassembly cards. ### Intro We saw this case getting made recently. As in this literal, exact case. _Editor's note: This was originally published on October 15, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication._ * * * #### Credits * * * ##### Test Lead, Host, Writing Steve Burke ##### Testing, Writing Patrick Lathan ##### Camera, Video Editing Vitalii Makhnovets ##### Video Editing Tim Phetdara ##### Camera Andrew Coleman ##### Writing, Web Editing Jimmy Thang * * * It was assembled in front of us on this assembly line when we toured the factory in China. While we were there, we asked Cooler Master if we could pick one from the line and then sign across the packing seals and panels to make absolutely certain it wasn’t tampered with. That’s kind of cool. Maybe this will kickstart the new way to ensure no review sample tampering. It’s kind of high effort and Cooler Master thought we were joking when we first pitched going to their factory to hand-pick a review unit, but they agreed. * * This case is interesting: It’s designed with a lot of mechanical freedom in mind for the user, meaning support for nearly any fan size up to 200mm. * * * * * * * Using its rails system with 1cm spacing. It also has support to invert the case, keep standard orientation but move the power supply to the top, flip the I/O to the top of the case, and use the 1cm-spaced holes for a variety of rails, tabs, fans, and cages. The case is built of sturdy, high-quality aluminum and steel components everywhere. * * * * It’s assembled of just corners and columns, which allows modularity; although, it was a letdown that the chassis itself is riveted together at the corners in a way that’s restrictive. The build quality is excellent and, although (like any case) it falls short in some areas, we like where this case series is going. The MasterFrame had a disappointing launch in 2021 with the MasterFrame 700. This one is better. It’s built from modular parts like an Erector Set (or Meccano if you're British or don't like saying the word “Erector”), and they're intended to be extremely reconfigurable as part of the "FreeForm 2.0" project. The MasterFrame 500 is $150 on Amazon right now, though it does change price a lot. We saw it change price 3 times in one month and ranged from $150 to $190, which is a wide range. ### **MasterFrame 500 Mesh Specs** Color(s)| Black/Silver ---|--- Materials - Exterior| SGCC Steel, Aluminum Materials - Side panel (Right / Left)| Tempered Glass Dimensions (L x W x H)| 471 x 261 x 544mm (Incl. Protrusions), 471 x 244 x 514mm (Excl. Protrusions) Volume| 60.9 L (Excl. Protrusions) Motherboard Support| Mini-ITX, Mini-DTX, Micro-ATX, ATX , E-ATX*, SSI-CEB* ​(*Conditional Support) Power Supply Support| ATX PCI Expansion Slots| 8 3.5" Drive Bays| 0 2.5" / 3.5" Drive Bays (Combo)| 1 (Bottom) Preinstalled Drive Bays - 3.5" / 2.5" Drive Bay (or SSD) (remark Toolfree quantity)| 1x 2.5 or 1x 3.5 I/O Panel - USB Ports| 2x USB 3.2 Gen 1 Type A, 1x USB 3.2 Gen 2x2 Type C I/O Panel - Audio In / Out| 1x 3.5mm Headset Jack (Audio+Mic) Pre-installed Fans - Front| ARGB: 2x SickleFlow 200mm ARGB, Non-ARGB: 2x SickleFlow 200mm Pre-installed Fans - Rear| ARGB: 1x 120mm ARGB, Non-ARGB: 1x 120mm Fan Support - Front| 3x 120mm, 2x 140mm, 2x 180mm, 2x 200mm Fan Support - Top| 3x 120mm, 2x 140mm, 2x 180mm, 2x 200mm Fan Support - Rear| 1x 120mm Fan Support - Bottom| 1x 120mm Radiator Support - Front| 120 / 140 / 240 / 280 / 360mm Radiator Support - Top| 120 / 140 / 240 / 280 / 360mm Radiator Support - Rear| 120 mm Dust Filters| Mesh Only (Removable Filters) Clearance - CPU Cooler| 190 mm Clearance - PSU| 235 mm Clearance - Graphics Card| 390 mm Included Accessories| GPU Support Bracket Cable Routing - Behind Motherboard Tray| 43 mm Warranty| 2 years Size| Mid Tower Series| MasterFrame Series _Specs copied from manufacturer materials, please read review for our own measurements and opinions_ ### **The Build** * * As it ships, the MF500 Mesh has 2x 200mm front intake fans and 1x 120mm rear exhaust. Cooler Master has had both some of the most successful and some of the worst implementations of 200mm fans, but in recent years, all of their 200mm fan solutions we’ve tested have passed testing. These fans work great if they’re unobstructed: They don’t have much static pressure performance, but do well with open flow while maintaining lower noise levels. Other than the size of the front fans, it's a fully conventional layout: the PSU shroud is at the bottom, there’s a normal motherboard orientation, and it has typical airflow paths. * * * The manual describes two main alternate configurations, which are moving the PSU to the top of the case and inverting the case. * * There are smaller items that can be moved around, like the front I/O and the drive mount, but those are the most transformative options. Cooler Master’s future expansion on this series hopes to add even more modularity, but we don’t review future promises. We review what’s actually shipping. The current DIY-style MasterFrame cases feel closer to the QUBE 500 Flatpack than they do to the MasterFrame 700. And that’s good, because the MasterFrame 700 sort of sucked. * * * Unlike the Flatpack, though, the MF 500 ships preassembled, and the pillars and corners that make up the "Exo-Structure" of the case are riveted, which may come as a surprise if you've been following the marketing. Cooler Master informed us that its short-term goal is to work with partners to manufacture cases in dimensions that the partners want, and the mid-term goal is to allow end users themselves to order custom case dimensions directly from Cooler Master. This is doable since it’s all just aluminum rods and corners. Long-term, it's possible that Cooler Master will sell Exo-Structure pieces for home assembly, but we were told it depends on cost. Currently, the amount of rivets means even replacing them with screws would make it a nightmare -- something we witnessed first-hand when we saw a hand-assembled model get taken apart at Computex over the span of 10-20 minutes. But the ambitions are good and we see a clear pathway to making this happen. Cooler Master wants to eventually take custom orders to assemble in its new California-based facility. * * The MF 500 also reminds us of the old CaseLabs designs, which had a similar ring of holes around the perimeter. For a case with a modular DIY focus, starting with a CaseLabs comparison is a good sign for Cooler Master. Case reconfiguration can take a while if you take the James May approach. The manual describes a 17-step process for moving the PSU to the top of the case, and then a further 10 steps for inverting the whole case, not including the optional 7 steps for inverting the front I/O. It's not as bad as the Dark Base Pro 900 (read our review), but unbuilding the system, inverting the case for the first time, rebuilding, and then running tests took most of a work day. * * It would have sped things up a little if the case didn't use a mix of M4 and 6-32 screws that look similar at a glance. It would have also sped things up if we took the Jeremy Clarkson approach. One of the major upsides of the case design is that fan mounting options are flexible. * * * Fans can either be centered by using tabs that attach to the edges of the case, or installed off-center using a long rail. The rails are sized to fit the sides/bottom/top and the tabs fit the front/bottom/top. We would have liked a couple of extra tabs to work with. Cooler Master lists compatibility with 120, 140, 180, and 200mm fans, but because the holes along the case perimeter are spaced 1cm apart, the rails in particular are compatible with any fan size under 200mm. This is pretty cool and means, if you wanted, you could throw a weird size like a 150mm or 160mm fan in there. The various case elements are made from surprisingly thick metal; for example, the tabs are 2mm thick and the long rails are 3mm thick. The result of that is a high build quality with good rigidity of the structure. This case isn’t flimsy. * * * * Support for 3.5" and even 2.5" drives is starting to wane, but we were surprised to see that such a modular case offers exactly one drive mount. Two metal "multi-function bars" are included in the accessory kit, and these can be used to either mount one fan, one SSD, OR one HDD. * * Cooler Master might sell extras, and it's likely someone will come up with a 3D-printed alternative, but the case would have more DIY cred if it included just two more of the basic steel strips. The case is set up perfectly to support more drives with its rail system and it just doesn’t. * * As the sticker on the case and instructions in the manual state, panel removal is TOOL FREE. Cooler Master has made a compromise here; the panels don't screw down, but the magnets that hold them on are extremely strong to compensate. There are also no obvious handles on the panels, so you may actually find it easier to pry the panels off with a tool. This is better than cases without a securing screw and an easily removed panel, but it’d still be nice to have a screw point to secure the glass. Another point that could use work is the cable management space. * * * This is one of those cases where you have to remember to route cables under supports or else the side panel can't close. The cutouts that the power cables must be routed through aren't all lined up with each other, either, so it takes some indirect routing. The motherboard tray is structured so that when the PSU shroud is moved to its secondary location, you have to remove the end of the shroud and route cables through that path rather than behind the motherboard tray, making it significantly harder to hide cables. * * As a positive, the cable bar was shaped perfectly for our motherboard, sitting level with the SATA ports and USB header along the edge of the motherboard so that they're hidden (although some SATA ports were blocked on our motherboard). The cable bar sits on two rails that run the full length of the case, so it can be moved up, down, back, and forth to fit flush against various motherboards. There are stone and wood veneer front panel options, but this is the least exciting aspect of the case's modularity to us, and none of the alternate panels that we've seen are ventilated, so they'd completely seal the front of the case, and we would strongly advise against them. * * Finally, there's a solid GPU support included in the accessory kit that can either stand under the GPU or hang from the roof of the case if it's inverted. ### **MasterFrame 500 Mesh Thermals** Grab a GN15 Large Anti-Static Modmat to celebrate our 15th Anniversary and for a high-quality PC building work surface. The Modmat features useful PC building diagrams and is anti-static conductive. Purchases directly fund our work! (or consider a direct donation or a Patreon contribution!) Time to get into thermal and acoustic benchmarking. Thermal testing was done with a mix of configurations. First was a standard layout with the stock configuration, followed by a stock configuration with the motherboard shifted down and the power supply mounted up top, and finally followed by the inverted configuration. We moved the power supply as part of the inversion, as that’s the intended use in the manual. In all configurations, we kept the PSU fan oriented to pull air from the exterior of the case (facing away from the PSU shroud). The GPU support was not installed for any thermal tests. The direct head-to-head competitor for the MasterFrame 500 Mesh is the Corsair Frame 5000D RS: it has the same focus on modularity, and it's roughly the same price as well ($190). The Frame 5000D RS isn't on our charts yet, so we'll make do with the Frame 4000D RS (read our review) with the caveat that it's smaller and significantly cheaper. Otherwise, the MF 500's price puts it up against the Fractal Torrent (watch our review), some variants of the Meshify 3 (read our review), the Antec Flux Pro (watch our review), and Arctic's brand new Xtender, although none of those cases have the same DIY marketing that the Corsair and Cooler Master cases do. #### **CPU Full Load Thermals - Noise-Normalized** We'll start with noise normalized testing, where we step down case fan speeds inside our hemi-anechoic chamber until we hit our 27 dBA SPL target, measured one meter from the front of the case. We published a video showing our process to build this chamber previously and have gotten regular use out of it since it was built. This allows us to control our acoustic environment so that outside noise doesn’t affect our results. Test hardware like this is made possible through your support on store.gamersnexus.net. The MF 500 landed in the middle of the chart at 46 degrees Celsius above ambient averaged across all cores and 49 degrees on the P-Cores. The Frame 4000D RS was cooler at 43 degrees all core, and the mesh-fronted cases closer to the MF 500's original $180-$190 price did even better, with the Torrent at 42, Flux Pro at 41, and Meshify 3 at 40. That's not bad, but it's also not impressive given that we'd expect the 200mm stock fans to shine in noise-normalized testing specifically since they’re quieter. $150 to $200 can get you a lot of performance from other manufacturers, and Cooler Master may be hoping that the modularity of the MF 500 will help balance things out. #### **GPU Full Load Thermals - Noise-Normalized** Moving to GPU thermals in the same test, the average GPU temperature in the MF 500 was 48 degrees Celsius above ambient and 53 degrees on the VRAM. That's behind the competition again, with the Frame 4000D RS averaging 45 degrees above ambient, the Meshify 3 averaging 43, the Flux Pro 41, and the Torrent 40. The MF500 is also beaten by the normal, cheaper Antec Flux. These results could be altered by tweaking the layout of the MF 500 in a number of ways, though, so let's cover the other configurations next. #### **CPU Full Load Thermals - Full Speed** With all case fans at full speed, the all core CPU average dropped to 42 degrees above ambient and the P-Core average dropped to 46. Moving the PSU to the top of the case dropped both averages by about two degrees, which is actually a noteworthy improvement and beyond just error. This change would also affect flow-through behavior, which we’ll look at in the GPU chart. Flipping the entire case in addition to that didn't cause any further change. Moving the PSU to the other side of the case means moving the PSU shroud and the motherboard tray as well, which alters airflow patterns and changes how the CPU cooler lines up with the front intake fans. For the given CPU performance, the MasterFrame 500 at 39.4 dBA is louder than its immediate neighbors on the chart, like the Lancool 215 (read our review) (which also uses 200mm fans) at 31.1 dBA. The Frame 4000D RS and Torrent are both louder and cooler, which is why we focus on noise-normalized tests for comparing cases, but the Flux Pro averaged 38 degrees all core and the Meshify 3 averaged 39 at roughly equivalent noise levels to the MF 500. Broadly speaking, the peak performance of the MF 500 using stock fans alone is underwhelming. #### **GPU Full Load Thermals - Full Speed** GPU thermals tell a different story: the stock case configuration averaged 47 degrees above ambient, but top-mounting the PSU caused no change, while inverting the case increased the average to 48. This isn't much of a change, though: we expect that moving or adding fans would have a greater effect than either of the case reconfigurations did. The GPU thermal results are weaker in comparison to the rest of the chart than the CPU thermals were, with the Frame 4000D RS at 42 degrees for average GPU temperature, the Meshify 3 at 41, the Torrent at 39, and the Flux Pro at 38. #### **VRM & RAM Full Load Thermals - Full Speed** We typically examine VRM and RAM thermals from the noise-normalized results, but in this instance it's more interesting to compare the different case configurations (which were tested with the case fans at full speed). The inverted case was best for VRM thermals with an average of 26 degrees Celsius above ambient, with the stock configuration and the top-mounted PSU configuration tied at 27. The inverted configuration was also best for RAM thermals with a 19 degree average, but the top-mount PSU layout beat the stock layout by one degree at 20 versus 21. ### **MasterFrame 500 Mesh Conclusion** Visit our Patreon page to contribute a few dollars toward this website's operation (or consider a direct donation or buying something from our GN Store!) Additionally, when you purchase through links to retailers on our site, we may earn a small affiliate commission. The MasterFrame 500 is a promising beginning for the new MasterFrame lineup. Cooler Master seems to go through cycles similar to Corsair where it executes excellently and then goes silent for a while or fumbles. We think the current series is well-built and has good panel build quality, has some good ambitions for its modularity, and is generally a well-executed but simple case. It’s sort of a utilitarian opposite to the more fanciful innovation Hyte is doing in its panel design or Fractal is doing with woods. This is a case that'll be at its best in the hands of modders that want to cut, print, and drill their own parts, which isn't something that gets shown in by-the-books testing. That said, we wish some aspects were more modular: specifically, it'd be cool to be able to disassemble the corners and edges that make up the basic structure of the case. We'd also like to have some spare multi-function bars, fan mounting tabs, and maybe some of the longer metal rails as well. The direct competitor for this case is the Corsair Frame 5000D (watch our review). Modular parts for the new Corsair Frame cases still haven't really appeared yet, although Corsair has offered us a motherboard tray for testing, so the two cases are in a race to see which will have useful optional parts available first. Cooler Master is already publicly hosting 3D files for the major components of the MasterFrame 500 Mesh (and MasterFrame 600) and it maintains a Printables page for accessories. Corsair began offering some 3D printing files during Computex but is overall behind on this aspect. The MF 500 also has heavier, more solid-feeling construction than Corsair's Frame cases, if that's something that appeals to you. In terms of thermals, there are easily higher-performance cases available for $200. Each of the high-end mesh fronted cases we compared to (namely the Meshify 3, Torrent, and Flux Pro) outperformed the MF 500, and the smaller, cheaper, and theoretically modular Corsair Frame 4000D RS ARGB (read our review) also generally performed better, but it is smaller. With the MF 500, you're paying for the ability to easily customize it (and the sturdy construction). Ultimately, at $150, this is significantly more competitive than at its $180 price despite its thermal deficiency versus other cases. It performs okay thermally in our benchmarks, but it isn't anywhere near as good as several other $180 and $200 cases, which is where this was priced when we started reviewing it. Fortunately, if it stays at $150, it drops to a lower price class, and that compensates a lot for the middling thermal performance. Also, at $150, we think the build quality is atypically good. Now, at $180 or $190, it's still good build quality. It's just there's other competition at that price point since you're getting into the kind of that upper echelon of cases. It's tough to nail down a recommendation since we don't know why the price keeps changing or where it'll stay. But if it stays at $150 and you're okay with having something that's not the best thermal performer but has this sort of modular mechanical approach to things that it's definitely worth considering. At $180 is still worth considering, but it's also worth considering a lot of the other competition that we named going through the thermal section. * * *

NVIDIA RTX PRO 6000 Blackwell Benchmarks & Tear-Down | Thermals, Gaming, LLM, & Acoustic Testsjimmy_thang September 25, 2025 ## We benchmarked the RTX PRO 6000 in gaming scenarios using an AMD Ryzen 9800X3D, experimented with LLM benchmarks, and ran thermal and acoustic tests The Highlights * The RTX 6000 features 24,064 CUDA cores, which is nearly an 11% increase over the RTX 5090 * Unlike the 5090, which uses liquid metal, the RTX Pro 6000 uses thermal paste * On the gaming side, the RTX Pro 6000 outperformed the 5090 by roughly 5 to 14% in our tests, but that alone isn’t a good reason to buy the card * Original MSRP: $8,000 - $11,000 * Release Date: March 18, 2025 #### Table of Contents * AutoTOC Grab a GN Tear-Down Toolkit to support our AD-FREE reviews and IN-DEPTH testing while also getting a high-quality, **highly portable 10-piece toolkit** that was custom designed for use with video cards for repasting and water block installation. Includes a portable roll bag, hook hangers for pegboards, a storage compartment, and instructional GPU disassembly cards. ### Intro We paid $8,000 for an RTX PRO 6000 GPU with 96GB of VRAM when we bought it from Brent Rambo, the actual guy in this meme, who’s grown up now and had access to an RTX Pro 6000. These days, Rambo is busy with custom and bespoke PC builds these days on his own website, and that means he has access to interesting hardware. It’s a shame that NVIDIA has left us feeling like it’s holding its engineers hostage for manipulating reviews, but it’s clear that we don’t mind buying our own hardware to review. _Editor's note: This was originally published on June 24, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication._ * * * #### Credits * * * ##### Test Lead, Host, Writing, Editing Steve Burke ##### Testing Patrick Lathan ##### Video Editing Vitalii Makhnovets ##### Camera, Video Editing Tim Phetdara ##### Testing, Writing, Camera Tannen Williams ##### Writing, Web Editing Jimmy Thang * * * * * The RTX PRO 6000 is the best marketing for AMD’s Ryzen 9800X3D (read our review) that we use in our GPU test benches, mostly because we’re seeing 5% to 14% scaling in gaming benchmarks versus the RTX 5090 (read our review). We also tore it down and found the most insanely dense PCB we’ve ever seen with a reversion away from liquid metal. But this doesn’t have 96GB of VRAM just because that’s where it all went when they lost it on the RTX 5080 (read our review), 5070 Ti (read our review), 5070 (read our review), 5060 Ti 8GB (read our review), and 5060 (read our review)-- it’s also there because it’s better for... AI. Today, we’re benchmarking some of those in a round of experimental benchmarks. We haven’t done machine learning and LLM tests before and we don’t have a full methodology defined yet, but we’ve begun experimenting. ### **NVIDIA RTX 6000 Overview** This card is not for gaming. That much is obvious. We’re still testing it in games, and that’s mostly for our own purposes: We want to know how much room the 9800X3D has in it to scale for GPUs, and fortunately, it looks like there’s still plenty for another generation of reviews. The 9800X3D keeps up with the RTX PRO 6000 even when at 1080p in some situations. The biggest difference between the PRO 6000 and the 5090 after the 96GB vs. 32GB memory capacity is the GPU itself: The PRO 6000 has 24,064 CUDA cores to the 21,760 of the 5090, nearly an 11% increase. That’s a big difference. There are multiple versions of the RTX 6000. The one we're looking at specifically is called the Blackwell Workstation Edition, but NVIDIA also has so-called “server editions” of the cards, and those are intended to go into servers and racks where they're getting force fed air from the front of the chassis rather than using an FE style cooler, which is the one we're looking at today. So, the one we're looking at is more similar to what we would see in sort of the consumer to prosumer workstation class. #### **NVIDIA RTX 6000 Specs, Architecture Basics, and Price** Getting into the specs: * * The RTX PRO 6000 Blackwell GPU is built on the GB202 Blackwell die. The full GB202 die has support for up to 24,576 CUDA cores, 768 TMUs, 192 ROPs that might even be present, and is also used for the RTX 5090. The PRO 6000 variation on this GPU has 24,064 CUDA cores, so it’s down by 512 and isn’t a perfect die, with 752 TMUs (down from 768) and still 192 ROPs, and they’d better all be there for the price. The card is also running 188 RT cores and supports PCIe 5.0 x16. Advertised clocks are 1590 MHz base and 2617 MHz boost. The real reason people buy this card is its memory, though, at 96GB of GDDR7. The closest consumer class card to compare against is the RTX 5090. It’s built on the same die, but with 176 ROPs, 680 TMUs, 170 RT cores, 21,760 CUDA cores, 32GB of GDDR7 memory, a 512-bit bus width, and a far lower price at $2,000 to, realistically, $3,000 rather than $8,000 to $11,000. For some notable differences: The 5090 only has a third of the memory capacity of the PRO card, and the RTX PRO 6000’s die is almost 98% present while the RTX 5090’s die is closer to 88.5% present, with the remainder either being fused off or defective fallout and fused off. You could also technically buy four 5090s for the price of one RTX PRO 6000, so the use case really does come down to having as much memory as possible attached to a single GPU and PCIe slot. ### **NVIDIA RTX 6000 Tear-Down** * * The RTX Pro 6000 we tested is a 2-slot card with 2 large fans on the front. The PCB is in the center of the card and you can see where the fins start to get taller. This indicates where the PCB plate ends. The card’s PCIe slot is connected via a customized pinout adapter. If you look at the backside of the card where the fans are, you’ll notice a slight depression down towards the middle of the fins. This is supposed to help with pressure drop. The design is full flow-through, which pushes air straight through the card rather than into a PCB. The other side of the card looks very similar to a 5090. The top of the RTX Pro 6000 retains the angle on the card’s vented slats. This projects the air out at an angle that’s up and away from the card. Like the 5090, the card uses a 12VHPWR cable. * * * Starting disassembly, we removed the card’s back plate. Then we removed a single screw, which allowed us to pull off the card’s exterior frame. * * * * From here, we removed some screws around the center of the card to remove the center covers. Doing this reminded us of how masterful the mechanics of the FE design is. It’s assembled really well, it’s easy to take apart, and there’s no mechanisms to try and prevent disassembly. * * * Next, we unscrewed the clamp that holds the PCIe slot to the card. Here’s a close look at the removed PCIe adapter. * * From here, we pried a plate off of the card, exposing the PCB. This gave us our first look at where the memory is connected. We counted 16 memory modules on the back with the clay type of thermal pads. * * * * The PCB is crazy dense, which is an engineering challenge. We then proceeded to remove ribbon cables and screws to the leaf spring to access the GPU. * * Pulling off the PCB, we were expecting to see liquid metal, but we saw thermal paste instead. We really weren’t expecting that. We also saw a massive GPU substrate. Upon removing the PCB, we noticed that 1 pad went over the edge a bit. Analyzing the paste application pattern, it looked like the heaviest imprint was right in the center with lighter pressure on the outer edges, which is fine. * * * Looking at the thermal pads, they were making clear contact, though we did see some pads that were smashed over the sides a bit. This is fine, however. We counted 32 total memory locations, which means that they are 3GB memory modules as that totals 96GB, the capacity of the card. Taking a closer look, we noticed some areas of poor contact along one edge of the inductors, where we noticed that half of the inductors are actually covered by a thermal pad. One of them only had about a third of the inductor covered. That’s not good as they really need to get the pad placement correct on GPUs that cost 8-11 thousand dollars. The board also has lots of little tiny components, leading to a very dense PCB. * * Finally, we cleaned off the thermal paste off of the GPU, exposing the GB 202-870-A1 SKU. This is a 5090 die, just a fuller version of it. ### **NVIDIA RTX 6000 Thermal Benchmarks** Grab a GN15 Large Anti-Static Modmat to celebrate our 15th Anniversary and for a high-quality PC building work surface. The Modmat features useful PC building diagrams and is anti-static conductive. Purchases directly fund our work! (or consider a direct donation or a Patreon contribution!) Thermal testing is up now. The RTX 5090 uses liquid metal with a 2-slot FE cooler that, genuinely, is one of the most impressive GPU coolers we’ve ever worked on. The prototype 4090 cooler we disassembled was more impressive in mechanical and cooling capabilities, but it was huge and impractical to manufacture. The 2-slot FE has done well in most cases, with the exception being memory cooling. We found that memory ran hot on our 5090 FE. That’s more of a concern with the PRO 6000 as well since it has so much memory. Here’s a thermal test during a fixed render workload. The 5090 ran at around 72 degrees Celsius for the GPU core temperature in this test, which is overall excellent considering its 2-slot design. The VRAM for the 5090 ran at around 90 degrees Celsius, which was hotter than we’re comfortable with. Once accounting for installation into a case and potential long-term implications of pad dryout and dust, that’s high. The PRO 6000 ran its GPU core at a significantly hotter 82 degrees Celsius in this test, so about 10 degrees warmer than the 5090. The GPU memory was about the same, measuring 88 degrees via software. #### **Fan Response** None of that means anything without fan speeds. Using only the auto settings, so following whatever VBIOS has programmed for the target GPU temperature, the RTX 5090 ran at 1550-1600 RPM for the average fan speed. The PRO 6000 ran at 1700. That’s a little faster with a much warmer core, but similar memory thermals. #### **Acoustics** Acoustics are next. For this testing, we’re using our hemi-anechoic chamber that we heavily invested in for improving our test quality. Our next major improvement will be upgrading the microphone, which is currently our bottleneck. We could hit a 6-8 dBA noise floor with a better mic, but we’re currently at 13.6 to 15.0 dBA for the floor. That’s still great though. The RTX 5090 ran at about 32.5 dBA, with the PRO 6000 at the same noise level; however, the noise floor was slightly different between these two, so adjusting for that, the PRO 6000 would be about 1 dBA louder under the same conditions. This is hardly noticeable as a difference, if noticeable at all for most people. Overall, the cooler design remains good when taking its size into consideration. The 5090 had higher noise levels in the higher frequency range, particularly 8000 to 10000 Hz. It also had a peak around 175 Hz, with the PRO peaking at 194 Hz and following a similar pattern, just adjusted right. Generally speaking, this mostly follows the same trend. Our frequency cutoff is 150 Hz, so the higher result on the left can be ignored. ### **NVIDIA RTX 6000 Gaming Benchmarks** Gaming benchmarks are next. #### **Dragon’s Dogma 2 - 4K** Dragon’s Dogma 2 is first. Tested at 4K, the RTX PRO 6000 landed at 140 FPS AVG, which has it ahead of the RTX 5090 by 5.8% for average framerate. Lows are not notably different. In other words, that’s about an extra $1,000 per 1% improvement, or about $780 per 1 FPS increase over the 5090 when calculating by the difference in cost for these two. The lead in the RTX PRO 6000 is about 42% over the 4090 (watch our review), with the 5090 already ahead of the 4090 by about 34% here. #### **Dragon’s Dogma 2 - 1440p** At 1440p, the lead over the 5090 is expanded to 6.3%, ensuring we get our full gaming value out of the card. Low performance is not different outside of the usual scaling along with the improvement in average frametimes. The RTX PRO 6000 hits 201 FPS AVG here, leading the 189 FPS of the 5090 and the 155 of the 4090. NVIDIA currently holds the entire top cluster of this chart, with AMD mostly focusing on the modern mid-range and Intel focused on the modern low-end. #### **Dragon’s Dogma 2 - 1080p** We next tested 1080p, which is what we all know buyers of this card really want to use it for. So-called AI use cases are obviously secondary compared to playing video games at 1080p. The RTX PRO 6000 leads this extremely important and pivotal chart with a 226 FPS AVG, dropping to just a 5.3% lead. Honestly, there is something valuable that comes out of this chart: It’s marketing for AMD’s Ryzen 7 9800X3D, which is somehow keeping up with these GPUs enough that it can still produce distinguished results between the $2,000-$3,000 5090 and the $8,000 to $11,000 PRO 6000. That’s impressive. This is more amusing to us as it proves the longevity of our bench hardware more than anything else. #### **FFXIV 4K** Final Fantasy 14 is up now, tested first at 4K. The RTX Pro 6000 ran at 208 FPS AVG, leading the 5090’s 184 FPS AVG result by about 13%. That’s a larger gain than we’ve seen in some of the other games, though obviously this is still a card intended for VRAM-intensive use cases and not gaming. That there’s still some power left beyond the 5090 is what’s more interesting, alongside the fact that the 9800X3D is so capable. Lows scale with the average and are not meaningfully better with the PRO 6000. #### **FFXIV 1440p** Tested at 1440p, we see a 10% improvement in framerate against the 5090, up at 348 FPS AVG for the 6000. Again, we’re mostly excited to see the 9800X3D continuing to scale and illustrating that there’s room left for another generation of testing on these benches. Lows remain in-step with the average. #### **FFXIV 1080p** At 1080p, the PRO 6000 again outperforms the 5090. This time, it’s reduced to about 4.3%. Let’s move on. #### **Starfield - 4K** Up next, we’re testing a card no one buys for gaming with a game no one plays. In Starfield at 4K, the PRO 6000 ran at 115 FPS AVG, leading the 5090 by about 7%. Lows are mostly within error. The value here is somewhere in the range of $700 to $1,000 per 1% improvement, depending on the price of the 5090. #### **Starfield - 1440p** At 1440p, the RTX PRO 6000 ran at 154 FPS AVG, leading the 147-148 FPS result of the 5090 marginally. The improvement over the 4090 is 17%, with the 5090 having previously led the 4090 by 12% in this test, as we’re becoming bound elsewhere. The 1% lows for the 6000 are noteworthy here: At 89 FPS for the averaged 1% low, the PRO 6000 outperforms the average framerate of the 5070 and nearly matches the 4070 Ti with just the PRO’s 1% numbers alone. #### **Resident Evil 4 - 4K** Resident Evil at 4K is up next. In this one, the RTX PRO 6000 ran at 219 FPS AVG, leading the 207 FPS AVG result of the 5090 by 6%. That’s consistent with most of the other tests so far. The lows, again, are not meaningfully different. The 4090’s 151 FPS AVG allows the 6000 a lead of 45%, or the 5090 a lead of 36.7%. #### **Resident Evil 4 - 1440p** At 1440p, the Blackwell workstation card ran at 371 FPS AVG, pushing the framerate measurably higher than the roughly 350 FPS AVG of the 5090 previously. We’re looking at about a 6.5% improvement, aligning with prior results again, and yet again showing just how good the 9800X3D is. #### **Black Myth: Wukong - 4K** Black Myth: Wukong is up now, tested first at 4K and representing one of our heaviest non-RT workloads. The RTX PRO 6000 didn’t break the 100 FPS barrier, but it was the first card to exceed 90 -- not that this particularly matters to anyone when the 5090 was already in the mid 80s. The 6000 outperforms our 5090 FE result by 7.4%, above what we’ve seen on average thus far by about 1 percentage point. 92 FPS AVG is a good showing in this benchmark when at 4K. The next card in the stack is the 4090 at 67 FPS AVG, then the 5080 at 58 FPS AVG. #### **Black Myth: Wukong - 1440p** At 1440p, the PRO 6000 ran at 138 FPS AVG and led the 5090 marginally, but it was technically still a measurable improvement. Total uplift is 6.1%, aligning with Resident Evil 4 previously. Once again, the next card is the 4090. After this is the 5080, which isn’t that different from the 4080 Super (read our review), which isn’t that different from the 4080 (watch our review), which isn’t that different from the 5070 Ti. #### **Black Myth: Wukong - 1080p** Black Myth: Wukong at 1080p is next. Selfishly, 1080p is interesting once again for probing at if there’s room in the CPU for more GPU. Turns out, there is: With the 9800X3D, the RTX PRO 6000 ran at 168 FPS AVG, improving on the 5090 by about 5%. That’s less than we saw at 1440p, which was less than we saw at 4K, so the card is gaining ground as resolution increases. #### **Dying Light 2 - 4K** Dying Light 2 is up next. At 4K, the PRO 6000 ran at 144 FPS AVG, which is a noteworthy improvement on the 126 FPS AVG of the 5090 FE. That’s a surprising 13.8%, which so far matches only one other game we’ve tested -- and that was Final Fantasy. Most of the other games are closer to 5-7%. The 4090 is next at 91 FPS AVG, meaning the PRO 6000 is 59% ahead of the 4090. #### **Dying Light 2 - 1440p** At 1440p, there’s still scaling: The PRO 6000 is now at 246 FPS AVG, leading the 5090 by almost 13%. That’s a drop from what we saw at 4K, but still more than most other games. #### **Cyberpunk 2077: Phantom Liberty - 4K** Cyberpunk 2077: Phantom Liberty is one of our next most intensive games after Black Myth: Wukong. At 4K, the RTX PRO 6000 breaks through 100 FPS and hits 108, leading the 5090 by 13.5%. This is actually a pretty big lead and is another one of the games that has a larger average improvement. That’s still not worth $8,000 or $11,000 or whatever it is they’re selling these for, but it’s interesting to see any difference at all given the focus on so-called AI workloads. #### **Cyberpunk 2077: Phantom Liberty - 1440p** At 1440p, the PRO 6000 ran at nearly 200 FPS AVG, about 9.4% ahead of the 5090. This remains one of the larger gaps for 1440p, although again, it’s obviously not something anyone should remotely consider primarily for gaming. The 4090 is down at 135 FPS AVG, meaning the PRO 6000 leads it by 48%. #### **Cyberpunk 2077: Phantom Liberty - 1080p** We’re leaving 1080p in because it’s the only place we see a bottleneck. Finally, we run into what seems like a CPU limit. The PRO 6000 and 5090 are at about the same level of performance, as is the 4090. Everything here is limited. That means our 1440p results were at least somewhat limited as well, with minimally the upper bound bouncing off of limits. ### **NVIDIA RTX 6000 Ray Tracing Benchmarks** Ray tracing benchmarks are next. #### **Ray Tracing - Black Myth: Wukong 4K** We’ll start with the heaviest one, which is Black Myth: Wukong at 4K with upscaling. In this test, the RTX PRO 6000 ran at 92 FPS AVG, which was barely any different from the 88 FPS observed on the 5090. We’re under 5% of uplift here. It’ll be interesting to see if RT workloads don’t produce as much change as rasterization. #### **Ray Tracing - Black Myth: Wukong 1440p** At 1440p upscaled, the RTX PRO 6000 ran at 131 FPS AVG, leading the 5090 by just 3%. There’s just not much change in this test in general. We’ll skip 1080p given the lack of improvement. #### **Ray Tracing - Cyberpunk 2077 (4K, RT Medium)** Cyberpunk at 4K with RT Medium is next. In this test, the RTX PRO 6000 ran at 65 FPS AVG, leading the 5090 by 11%. That’s roughly in-line with what we saw in Cyberpunk rasterized. This is without any upscaling, which may be helping the PRO card’s relative gain. And finally, unlike what we saw in the 5060 and 5070 class cards with significant VRAM limitations harming low performance, the PRO doesn’t have any VRAM issues. This is clearly the solution. The solution is to buy $8K-$11K video cards. As before, note that some cards with low VRAM, like the 5070 and 4070 Ti, have averages that look better than the reality (even though the averages are also not great). #### **Ray Tracing - Cyberpunk 2077 (4K, RT Ultra)** 4K with RT Ultra is next. We almost never publish this test since it’s so intensive that it becomes somewhat useless, but it’s interesting here. The RTX PRO 6000 ran at 56.6 FPS AVG, leading the 5090 by 6.4%. That’s a significant drop from what we saw at RT Medium, indicating that the increase in intensity for the RT workload is minimizing the benefit of other aspects of the PRO 6000’s improvements. It is getting overrun by the RT workload intensity. The lead over the 4090 is about 44% here for the 6000, with the 5090 leading the 4090 by 35%. Note that the lower portion of this chart is unreliable for average framerate since the cards are all unplayable and stuttering. This can sometimes be due to exceeding VRAM limitations, such as on the RTX 5070. The average looks far better than the reality, even though the average is also unplayable. #### **Ray Tracing - Cyberpunk 2077 (1080p, RT Medium)** At 1080p with RT Medium, the PRO 6000 ran at 182 FPS AVG with lows paced proportionally. That has it ahead of the 5090 by about 12 FPS AVG, or about 7%. #### **Ray Tracing - Dragon’s Dogma 2 4K** Dragon’s Dogma 2 with ray tracing and without upscaling is next. Tested at 4K native, the RTX PRO 6000 at 121 FPS AVG leads the 5090 FE by about 7% in average framerate. This is starting to match a pattern, so we’ll let this be our last game test. Let’s move on to something else. ### RTX 6000 **AI Benchmark Charts:** Now, we'll get into some of our first ever LLM and machine learning benchmarks. As a disclaimer, we know enough to get some charts together to run some benchmarks where we feel like the controls of the test environment are good, but we do not yet know enough to have a full picture of capabilities outside of the test suite that we're using currently. We've tested with LM Studio, 3DMark testing, and ML Perf. We kind of settled on one set just to experiment for now. And this is our disclosure that we are experimenting with new testing, which means that the charts get the experimental chart label on top of them. And that's just so everyone's aware that these are not up to our full confidence standards of our normal benchmarks that we do all the time because we're still learning here. #### **LM Studio Testing** For our LM Studio testing, we recorded the response speeds from 8 different models that incrementally increase in size. We’ll get to larger models in a second. Our first chart illustrates the AVG tokens/second results from the three smaller models – ranging from 8.5GB to around 19GB in size. We think these first three are more representative of like-for-like tests compared to some other models we’ll be going over today – because any of the GPUs on our chart can load these models without being limited by VRAM capacity. In DeepSeek Llama 8B Distil, the RTX PRO 6000 is functionally tied for first place with the RTX 5090 FE – both with response speeds of roughly 81 tokens/second, improving on the 4090’s 62 tokens/second AVG by about 30%. In 8-bit Phi-4, the PRO 6000 puts some distance between its own tokens/second AVG of 62 and the 5090's 51 tokens/second AVG, showing a 22% improvement. Compared to the 4090, the RTX PRO improves by about 43%. In Qwen 2.5, the PRO 6000's 44 tokens/second AVG sees improvements of 25% over the 5090 and 37% over the 4090. #### **Text Generation** Moving to our next chart, we’ll really start to see performance gaps begin to widen as VRAM limitations take effect. The RTX 6000 tops the chart in all five models. In InternLM, the RTX PRO 6000's 50 tokens/second AVG improves upon the 5090's 40 by 25% and upon the standard 24GB 4090 by 319%. In Mistral Small 26GB, the 4090 and 5090 see much sharper performance decreases than the PRO 6000. The workstation GPU achieves 42.4 tokens/second AVG, or 147% improved from the 5090’s 17 and 560% greater than the 4090’s AVG of 6.4. Once we reach Gemma 3 27B, the workstation card really starts to separate itself from the lowly $2,000 gaming cards. In this model, the RTX PRO 6000 achieves a 29 tokens/second AVG, with the 5090 only reaching 5 and the 4090 only seeing 4 tokens/second, give or take some change. The comparisons remain largely the same in QwQ 32B. In the final Llama 3.3 70b Q4_K_S model we tested, the RTX PRO 6000 achieved its greatest lead yet. The workstation card sees improvements of 928% over the 5090 and 1141% over the standard 4090. A more simplistic way to interpret this data is: At a certain point, the other cards run out of VRAM, while the workstation GPU doesn’t even use half of its VRAM in the largest Llama model we tested. You could expand beyond this as well, of course, but that’s the extent of our experimentation for now. ### **NVIDIA RTX 6000 Conclusion** Visit our Patreon page to contribute a few dollars toward this website's operation (or consider a direct donation or buying something from our GN Store!) Additionally, when you purchase through links to retailers on our site, we may earn a small affiliate commission. That's it for the RTX Pro 6000 test. It's a review in a way but at the same time, we'd want to do a lot more tests in the ML category to really have a fully-fledged review. This piece has been an exciting experiment for us. The biggest discovery for us was seeing no liquid metal while we were doing the tear-down. That was probably known, but it’s not something that we had looked into. So that was interesting. We could see why NVIDIA did that for reliability reasons. We could also see it for liability reasons if the company is worried about liquid metal leaking out onto expensive servers. We kind of doubt this was done to save on costs, though maybe it was done for liability costs, but it just doesn't seem like NVIDIA is going to try and save a buck on this kind of card. We spoke to Wendell from Level1Techs and he noted that there's some apparent buggy behavior with Blackwell right now and so that's still getting updated and perhaps there's performance optimizations that could be done. On the side that we're familiar with, which is the gaming aspect, seeing 5 to 14% improvement is mostly useful not because you should ever in any way consider this card for gaming, because that would be an insane waste of money if that's all you would do with it, but more because it shows us that there's room left in the 9800X3D for scaling, which is pretty cool. * * *

Fractal Epoch Gaming Case Review | Thermals, Cable Management, & Build Qualityjimmy_thang September 23, 2025 ## We analyze the Fractal Epoch PC case against some of the best computer cases on the market. The Highlights * Fractal's Epoch computer case builds upon the Fractal North chassis, but swaps the front panel to accommodate larger GPUs while driving the price down * At its price range, we’d buy something else * The Epoch aims to be more price competitive than Fractal's other modern offerings, helping breathe life into the price category that the Pop Air has occupied * Original MSRP: $110 MSRP ($130 for RGB) * Release Date: September 2025 #### Table of Contents * AutoTOC Grab a GN Tear-Down Toolkit to support our AD-FREE reviews and IN-DEPTH testing while also getting a high-quality, **highly portable 10-piece toolkit** that was custom designed for use with video cards for repasting and water block installation. Includes a portable roll bag, hook hangers for pegboards, a storage compartment, and instructional GPU disassembly cards. ### Intro The Fractal Epoch is the Fractal North (watch our review) with a new front panel and a lower price. That makes this simple, except for one massive problem: The name. According to Google, this is pronounced the Epic. According to the Cambridge dictionary, it’s either the E-poch, eh-pock, or epic. _Editor's note: This was originally published on September 2, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication._ * * * #### Credits * * * ##### Test Lead, Host, Writing Steve Burke ##### Testing, Writing Patrick Lathan ##### Camera, Video Editing Vitalii Makhnovets ##### Camera Tim Phetdara ##### Writing, Web Editing Jimmy Thang * * * * * * * * * * * Fractal's newest case reuses a ton of tooling from the popular North in order to keep costs low, resulting in a $110 MSRP ($130 for RGB). The Epoch's front panel increases GPU clearance without making the case any larger than the compact North it's based on. The new design also comes with 3 stock fans versus the North's 2, so we'll cover the differences in thermal performance in detail, as well as noise testing. The North itself originally launched at $130, then went to $140, and is now $154. This is almost definitely due to tariffs. With the Meshify 3 (read our review) going for $155 minimum as well, the Epoch fills a budget slot in Fractal's lineup that hasn't really been updated since the Pop Air (watch our review) and Focus 2 came out in 2022. Today then, we’re reviewing and benchmarking the Fractal Epoch for thermals and acoustics. 3.5"/2.5" drive mounts| 3 (2 included) ---|--- Dedicated 2.5" drive mounts| 2 5.25” drive mounts| 0 Expansion slots| 7 Motherboard compatibility| ATX / mATX / Mini-ITX Power supply type| ATX Front interface| 1x USB Type-C 20Gbps2x USB Type-A 5Gbps1x Audio/Mic Combo Jack Total fan mounts| 6 x 120 mm or 4 x 140 mm Front fan| 3 x 120/2x 140 mm(# 3 x Momentum 120mm PWM included) Top fan| 2 x 120/140 mm Rear fan| 1 x 120 mm Bottom fan| N/A Dust filters| PSU Front radiator| Up to 360mm; Up to 280mm Top radiator| Up to 240mm Rear radiator| 1x 120mm Bottom radiator| N/A PSU max length| 1 HDD Tray: 255mm max2 HDD Tray: 155mm max GPU max length| 372 mm with front fan mounted; up to 345mm with front mounted radiator. CPU cooler max height| 170 mm Cable routing space| 30 mm Cable routing grommets| Yes Fixed velcro straps| Yes Tool-less push-to-lock| Thumb screws for side panels, PSU bracket, ball joints front panel and sliding lock top mesh. Captive thumbscrews| HDD brackets, SSD brackets, Side panels, PSU bracket Left side panel| Tempered glass 3 mm [or] Steel Right side panel| Steel Case dimensions (LxWxH)| 447 x 215 x 469 mm Case dimensions w/o feet/protrusions/screws| 443,3 x 215 x 455,3 mm _Specs copied from manufacturer materials, please read review for our own measurements and opinions_ ### **The Build** The build experience in the Epoch is largely the same as it is in the North, but we'll recap some details here. We like the North overall, but at launch, we stated that the North felt like it should be $20-$40 cheaper if it weren't for the wood in the front panel. We also said the wood elevated the case and (arguably) made it worth the price. The Epoch lacks wood and it's $20-40 cheaper than the North is currently, so that lines up with the earlier statements. The only challenge is that it’s $30-$40 cheaper than the North’s new price, which is higher. * * * * The Epoch (like the North) is fairly basic. Removing the two-part side panel and the sliding mesh top panel gives good access to the case interior, but there's a clear divide between the Epoch and something like the more expensive Meshify 3, which has fully removable top and front mounts that completely open up the case. * * * * The North's shallow form factor means that GPU clearance is limited to 355mm, which would mash the end of the GPU directly against the front fan mounts. The front of the Epoch has been reworked to push the front fans further into the front panel, hence the listed 372mm GPU clearance. Updating this for the Epoch was worthwhile, although it seems more likely that people are going to try to fit big, expensive GPUs into the North or North XL rather than the plainer Epoch. * * * * The redesign also means that the new case has more clearance for fans and radiators between the PSU shroud and the fan bracket, up from 36mm to 58mm by our measurements, and there's less material overhanging the front drive bracket. The only downside we noticed with the front panel redesign is that fan cables need to be managed more carefully to avoid getting in the way. * * The new design is simpler, with the metal mesh at the front of the Epoch acting as a filter rather than the removable filter in the North. We’ve found that this is better thermally (and in the past, we’ve found in testing that noise-normalized, it’s also superior). You’ll just wipe the front panel down instead. * * Our main complaints with the North boiled down to cable management, which was lacking compared to competitors even at the time. That’s because the deeper channel at the front of the case is 3cm deep at most, which can lead to the side panel bulging with poor cable management and a lot of cables, especially since the side panel doesn't hook into the case along that edge. It's not an insurmountable problem, which is good, because the Fractal Epoch is exactly the same in that regard. * * * The cutout for CPU power connectors is also a little on the small side, and the interior of the PSU shroud can get crowded if both drive trays are used. These space issues are a little more forgivable in the North because the North XL exists as an alternative, but we aren't aware of plans for an Epoch XL. * * * * A couple of features from the North have been removed, presumably to cut costs. There's currently no mesh side panel option for the Epoch, which also means that there's no side fan bracket (and no mounting holes to install one). There's also no built-in fan hub, but Fractal has handled this by daisy-chaining fans in both case variants. * * * As we saw in the Meshify 3, Fractal uses Type-C shaped connectors (NOT actual USB) for its RGB Momentum fans, but the RGB Epoch uses an adapter that splits out into standard fan and RGB connections. Otherwise, the differences are neutral cosmetic details. * * * * * The front I/O has been rearranged, the illuminated strip from the North is gone, and the two audio jacks have been combined into one. The USB ports are of the same type and quantity, but the Type-C port is now connected with a flat ribbon cable. The case feet have been redesigned without metal accents (the Epoch's might still be more expensive since each is a unique shape), and the faux-leather pull tab at the back of the case has been replaced with fabric. The captive screw on the half-panel at the front of the case still gets wedged, just like the North, but at this point, we've had to make peace with that. ### **Fractal Design Epoch Thermals** Visit our Patreon page to contribute a few dollars toward this website's operation (or consider a direct donation or buying something from our GN Store!) Additionally, when you purchase through links to retailers on our site, we may earn a small affiliate commission. For thermal testing, we ran one set of tests with the stock non-RGB case, then moved the RGB fans into that same case and ran another set of tests to ensure every other variable remained exactly the same. Of the cases we've tested lately, the direct comparisons for the Epoch are the Corsair Frame 4000D RS, Phanteks XT Pro Ultra, the Lian Li Lancool 207 (or arguably the 217), and of course Fractal's own North. These are all fairly compact ATX towers in the $90-$110 range, with the Epoch landing at the upper end of that and the North being the highest of all of them. The non-RGB Epoch will be our main point of comparison since the RGB version costs $130 and could have slightly worse thermal performance. According to the spec sheet, just like with the Meshify 3, the non-RGB fans can generate more noise and move more air at peak speed. Now we have an opportunity to actually test for differences, although they should be minor, and the rated RPM is 2200 for both models. #### **CPU Full Load Thermals - Noise-Normalized** We'll start with our noise-normalized test, where we adjust the speed of the stock case fans to hit an overall SPL of 27 dBA as measured in our hemi-anechoic chamber. We built this chamber to allow more accurate testing and normalization by eliminating external noises that may affect day-to-day noise floor. With a full-system torture workload, CPU temperature in the Epoch was 43 degrees Celsius above ambient for the all-core average and 47 degrees on the P-cores, and the temperatures with RGB fans were within one degree for both measurements for the 2 cases. Fractal's claim is that the non-RGB fans are a little louder and higher CFM, which could mean that our noise normalization reverses or equalizes for performance differences. The Epoch outperformed the original North, which came with two 140mm fans and averaged 46 degrees above ambient all-core and 50 degrees P-core. The glass-sided North is more similar to the Epoch, but the mesh option is an advantage for the case, lowering the all-core average to 44 degrees above ambient and 48 degrees over ambient P-core. The Phanteks XT Pro Ultra was even warmer than the glass North, which means the Epoch is also better than the XT Pro Ultra. The Corsair Frame 4000D RS was within error of the Epoch and functionally the same, which makes sense given the similar layout and fan arrangement. The most serious competition comes from Lian Li's dirt-cheap Lancool 207 (read our review) with its 41 degree average, as well as the more expensive 217 (read our review), also at 41. These cases are at about 45 degrees for the P cores. #### **GPU Full Load Thermals - Noise-Normalized** Moving to GPU thermals in the same test, the Epoch RGB averaged 45 degrees above ambient and the regular Epoch averaged 46. Memory temperatures showed a slightly wider delta, with the Epoch RGB at 49 and Epoch at 51. The regular Epoch was louder at max speed, so it required a greater fan speed reduction to hit our noise-normalized threshold, which explains the temperature difference in combination with a +/-1 degree range. The regular Epoch outperformed the glass North's 48-degree average and tied the mesh North at 46. The Phanteks XT Pro Ultra came closer to the Epoch here than it did for CPU thermals, although the memory temperature was a weak point for the Phanteks case at 54 degrees over ambient. The Frame 4000D did well with a 45 degree GPU average, while the Lian Li cases again are ahead, with the 207 and 217 "GPU Mode" showing similar performance at 42 degrees above ambient on the GPU. Credit to Fractal for still having the best GPU thermal performer with the Torrent (read our review) though, tying the Flux Pro (read our review). These are much more expensive cases. #### **GPU Full Load Thermals - Full Speed** We’re switching to full speed tests now, which means we allow noise to become a variable. The noise level is next to the case name. At full speed, the non-RGB and RGB cases didn't show much of a difference in thermal performance, but they did in noise. The stock Epoch with non-RGB fans measured 42.2 dBA, while the RGB case measured 39.3 dBA. That lines up with Fractal's claims, which have a regular Momentum 12 at 31.3 dBA and a Momentum 12 RGB at 28.03 dBA in their testing. These hard numbers are different because they are tested in different environments and methods -- we’re measuring total system noise, not just the fans. The average GPU temperature with the RGB fans in the Epoch was technically lower at 43.6 degrees average, but not enough to be outside the margin for error, and the average GPU memory temperatures were nearly identical between RGB and non-RGB. The North with its stock fans almost kept up, with the mesh-sided result only one degree warmer and the glass-sided result at 46 degrees above ambient, although noise levels were higher than the Epoch. Each of the other competitor cases we mentioned performed better, though, starting with the XT Pro Ultra at 43 degrees, then the Frame 4000D at 42, the Lancool 217 at 39-40 in an ideal configuration, and finally the Lancool 207 at 38 degrees above ambient and roughly the same noise level as the Epoch. Noise levels have the Epoch about the same as the Antec Flux Pro when at max fan speed, with the Flux Pro running far cooler with its superior fans. The Flux Pro is also $165 to $180, depending on promos, so it’s not the same price class. #### **GPU Full Load Thermals - Standardized Fans** Using our standardized set of two 140mm intake fans and one 120mm exhaust allows us to make a true head-to-head comparison between the North and Epoch front panels, without stock fans as a variable. In terms of GPU thermals, the Epoch, the mesh-sided North, and the glass-sided North all averaged the same: 46 degrees above ambient. The Epoch's front panel may be cheaper, but it's not worse for thermals. #### **CPU Full Load Thermals - Standardized Fans** CPU thermals between the two North variants in the standardized fan test were also very similar, but the Epoch was about one degree cooler, averaging 38 degrees above ambient all-core and 42 on the P-Cores alone. The new front panel seems to be slightly more open, but any further performance advantages for the Epoch are down to its stock fans. #### **VRM & RAM Full Load Thermals - Noise-Normalized** Back to the noise normalized test, the Epoch RGB's VRM temperature averaged 29 degrees above ambient and the regular Epoch averaged 30. That's better than either North variant, but 217 remains top of the chart at 26 degrees. ### **Fractal Design Epoch Conclusion** Grab a GN15 Large Anti-Static Modmat to celebrate our 15th Anniversary and for a high-quality PC building work surface. The Modmat features useful PC building diagrams and is anti-static conductive. Purchases directly fund our work! (or consider a direct donation or a Patreon contribution!) The Epoch is entering a market where every dollar counts. This is a harder market to compete in than the $150-$180 range. * * * We think $110 is at the upper limit of what Fractal can realistically ask for this model. It's a North without the wood, and while the trendsetting wood accents will keep the North selling even at $140+, a North without the wood is a reliable but unexciting compact mid-tower, even if the performance is a little better and there's a little more GPU clearance. Fractal has steep competition from Corsair, Phanteks, and Lian Li, among others. * * Corsair's Frame 4000D RS (the one with three stock fans) is basically the same thing as the Epoch, and as of this writing, it's on sale for $95. We weren't terribly impressed with the Phanteks XT Pro Ultra's thermal performance, but it's currently even cheaper at $80 after a rebate on Newegg. The most ruthless competition is from Lian Li, as usual, with its cutthroat prices. It's selling its wood-trimmed Lancool 217 for as little as $120, but a more direct comparison for the Epoch is the $82 compact Lancool 207, and even the 207 Digital with an integrated LCD is on sale for $105. It’s good that Fractal is updating its budget offerings because it hasn’t had a refresh here in a couple years, but there's not a particularly strong reason other than brand loyalty to choose the Epoch over its cheaper competition. The North is a good case, but it’s primarily good because of the front panel and the optional mesh side panel. The Epoch is the North without either of those, leaving it as simply “fine.” It’s OK. But competing in the lower price classes, we’d buy something else. * * *

AM4 Lives: AMD Ryzen 5 5500X3D CPU Review & Benchmarksjimmy_thang September 22, 2025 ## The 5500X3D likely takes fallout from the manufacturing process and then re-spins it as a new model, allowing AMD to salvage silicon The Highlights * The AMD Ryzen 5 5500X3D CPU is for the Latin America market and builds upon the AM4 X3D CPU lineage and is a 6-core, 12-thread part * The 5500X3D isn’t a particularly good performer in things like production applications, compression, and decompression * AMD is still launching AM4 CPUs, which is awesome since the socket is now nearly 10 years old * Original MSRP: $240-$250 (approximately) * Release Date: June 2025 #### Table of Contents * AutoTOC Grab a GN15 All-Over Print Component Mouse Mat for a high-quality mousing surface that'll fit your keyboard & mouse. These mouse mats use a high-quality yellow rubber underside, a blue stitched border for fray resistance, and are covered in PC parts. This is the best way to support our work and keeps us **ad-free** to support consumer-first reviews! ### Intro AM4 is the GOAT of motherboard sockets at this point. The socket launched in 2016 and was used for AMD’s first Ryzen CPUs in 2017, and AMD is still launching CPUs for it in 2025. The newest one is the Ryzen 5 5500X3D CPU, a 6-core, 12-thread part that follows other unexpected launches. AMD launched the R7 5700X3D (read our review) in 2024 and the R5 5600X3D (exclusive to Micro Center) in 2023. The R5 5500X3D has a lower frequency than the 5600X3D and is exclusive to Latin America. Our viewer Leo in Brazil helped us buy one to review, and it’s currently priced at about $204 converted to USD and launched at around $240-$250 USD a couple months ago. _Editor's note: This was originally published on August 29, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication._ * * * #### Credits * * * ##### Test Lead, Host, Writing Steve Burke ##### Testing Patrick Lathan ##### Testing, Video Editing Mike Gaglione ##### Camera Tim Phetdara ##### Writing, Web Editing Jimmy Thang * * * Today, we’re benchmarking the R5 5500X3D. We’re going to keep this set of benchmarks relatively simple and straight-forward: We’ll look primarily at gaming, frequency validation, and we’ll add some production benchmarks. We’re currently revamping and overhauling our efficiency testing as part of our regular bench suite update interval, so we won’t be running those numbers today. ### **Overview** We’ll start with a quick price comparison. #### **CPU Market Pricing Update** So again, the 5500X3D was about $240-$250 for our viewer, as seen in the screenshot above from the retailer who sold it. About a month or two later, it went down to $204. That’s still pretty expensive by US market standards, but we’re admittedly not very familiar with the Brazilian or South American markets where this is sold. Starting with some pricing updates from the market we do know in the US, here’s what it looks like: Currently, the Intel 265KF without IGP is $283 (apparently marked down from $300), the 9600X is $205, the 14600K (watch our review) is $190, the R7 7700 (watch our review) is $285, and the Ultra 5 225 from Intel is $212 (we haven’t tested this one). The 5800X3D (watch our review) is mostly gone from retailers in any reasonably priced capacity and the 5700X3D is also only available at higher prices and from third-party sellers now. To set the price for the high-end, a 9800X3D (read our review) is about $480. These are US prices though. Checking the same website our viewer used, we found the 5600X for about the same price as the current 5500X3D price, the Intel 245K for almost exactly twice the current price of the 5500X3D, and the AMD R5 9600X for $270. Given these prices, the 5500X3D at its current $204 is one of the better prices on this particular website. But again, don’t take us as experts for this space. We only learned of this site a month ago. #### **Specs & Price** The AMD R5 5500 non-X3D (watch our review) had a major downside from the R5 5600, which was the cache: The 5500 had just 16 MB of L3 Cache, down from 32 MB on the R5 5600. This was a huge change. Despite having the same core and thread count, same 65W TDP, and frequencies which are overall close enough to be comparable, this loss of cache materialized heavily in gaming performance right away in most cases. The 5500X3D has 96MB of L3 Cache, as we’ve come to expect from its class of CPU. It also has a 105W TDP, which increases its power budget; however, the 5500X3D’s clock speed is significantly lower than that of the 5600 (watch our review) and even 5500, at 3 GHz base and 4.0 GHz boost. This will be its biggest potential downside, but we’ve often seen that cache can make up for lower frequencies in gaming. If the CPU is a downbin or cobbled together from low-performing silicon, then these frequencies make sense as it might be comprised of chips that couldn’t hit higher targets. Likewise, it's possible if it's lower quality silicon that it could require more power to drive the frequencies it is maintaining. The rest of the specs are familiar: The 5500X3D uses a 7nm FinFET process from TSMC, like the 5500/5600, uses the AM4 socket that’s been GOATed at this point, and is a 6-core, 12-thread part. #### **Frequency Validation - All-Core** We’ll start with frequency validation in Blender rendering to ensure the CPU is working as advertised. AMD’s spec sheet defines a 4GHz max boost and 3GHz base clock. This all-core workload in Blender has the 5500X3D at 3950MHz, falling below the max advertised boost. This is expected for an all-core workload, as long as the CPU can hit the maximum advertised boost in a single-core workload. The non-X3D R5 5500 CPU ran at 4250MHz, so it has a frequency advantage. The 5600X3D held a frequency of 4350MHz, making it the fastest of these 3 CPUs by frequency. All 3 of these are 6-core, 12-thread CPUs. #### **Frequency Validation - Single-Core** This next chart shows the maximum single core frequency per interval across a Cinebench 1T benchmark. The 5500X3D had a maximum frequency of 3950MHz, which means it falls short of AMD’s target of 4000 MHz maximum advertised boost. Technically, the spec sheet says “up to,” but that’s bullshit. AMD has done well since the 3000 series to ensure that its CPUs hit this advertised frequency in one of these two workloads for every other CPU we’ve tested for years, making this the first to fail in a long time. 50 MHz short of the target is disappointing. There are a few blips later in the test as tiles change, but the bulk of the test does not meet this target. The 5500 non-X3D hit 4250 MHz again, with the 5600X3D at 4350 MHz. ### **5500X3D Gaming Benchmarks** Visit our Patreon page to contribute a few dollars toward this website's operation (or consider a direct donation or buying something from our GN Store!) Additionally, when you purchase through links to retailers on our site, we may earn a small affiliate commission. #### **Stellaris Simulation Time Benchmark** Stellaris simulation time is up first. This benchmark is CPU-intensive and also a gaming benchmark, but it’s useful for its real-world representation of something beyond framerate. This lets us look at the actual time required to simulate change in the game, meaning the impacts are felt in real time. X3D stacks up like this: The 5800X3D is at 50 seconds required, the 5700X3D is at 55 seconds, the 5600X3D required 55.6 seconds, and the 5500X3D required about 61 seconds. The best performer is the 9800X3D at 37.1 seconds, illustrating significant scaling headroom. Compared to the 5500X3D, the 5800X3D required 17% less time to complete the work. The 5700X3D required about 10% less time to complete the work, with the 5600X3D at about 8% required time reduced. Intel’s 12600K (watch our review) has about the same performance as the 5500X3D. The 12700KF is a bit better, with the 245K (read our review) improved somewhat notably to 52 seconds, though we still don’t recommend the 200 series of CPUs. #### **Dragon’s Dogma 2 Benchmarks** Dragon’s Dogma 2 is up next. This is a relatively CPU-intensive benchmark. The 5500X3D ran at 95 FPS AVG here, with lows well-paced and frametimes in step with the average frametime. This has it around the same level as the 12600K, matching our Stellaris lineup. The 9800X3D shows clear room for scaling at 132 FPS AVG, or a 40% improvement over the 5500X3D, but would also require a new platform with a higher cost. The 5800X3D shows the best gaming capabilities of the AM4 socket with its 108 FPS AVG result, a 14% uplift over the new X3D part. The 5700X3D is closer to the 5500X3D, with the 5600X3D outperforming it. This isn’t abnormal and we’ve explained this over the years: The 5600X3D has a higher 4.4GHz turbo clock, with the 5700X3D at 4.1GHz due to the same 105W power budget being spread across more cores. The 5500X3D is comparable to the 5700X3D, 245K, and 12600K in terms of performance. The 5700X3D still retains the benefit of more cores in situations where that’s useful. #### **Baldur’s Gate 3 Benchmarks** Baldur’s Gate 3 is up now. This one is lightweight on GPUs, but can be surprisingly good for showing CPU scaling. The 9950X3D (read our review), 9800X3D, and everything below them prove that. We’re still scaling all the way up to the best two CPUs on the market. The 5500X3D held a framerate of 102 FPS AVG; however, these are slightly worse than the lows from the neighboring two Intel CPUs. The 265K and 12700KF both are better in frametime consistency, despite the 5500X3D not being bad. The 5800X3D leads the 5500X3D by 19%. The 5600X3D and 5700X3D are similarly ahead of the 5500X3D, both at about 9% ahead of the new CPU. The 4GHz clock and 6 cores of the 5500X3D are limiting it here, despite overall fine performance. #### **Final Fantasy XIV: Dawntrail 1080p** Final Fantasy 14: Dawntrail is up next, tested at 1080p first. The 5500X3D ran at 297 FPS AVG, so nothing to be upset about. Frametime pacing is overall good. The 5700X3D is technically better, but not by much; its 302 FPS AVG is an uplift of 1.7%, with the 5600X3D benefitting from the higher frequency. We’ve seen this behavior in Final Fantasy for years now, where the higher frequency helps more than an extra 2 cores. The 9800X3D helps establish a ceiling around 370 FPS AVG, with the 9950X3D pushing close to 400 FPS. There’s room to scale here. The 5500X3D manages to outperform the 14900K (read our review) and Intel 285K (read our review). We’ve talked about this before, but Final Fantasy’s chart flipped from an update a couple years ago now: It used to be that Intel held the entire top half of the chart, and now it’s AMD. That remains true with the X3D AM4 parts while the non-X3D AM4 variants fall below some of the Intel CPUs. #### **FFXIV 1440p** 1440p is more or less the same, but we’re including it just to show scaling. Everything is a little bit truncated by the imposition of more GPU load, but particularly the 9800X3D and 9950X3D. These two CPUs are now about tied, with everything down to the 7600 (watch our review) at least occasionally glancing off of the GPU limit, even if rarely. #### **F1 24 1080p Benchmarks** F1 24 is next. In this one, the 5500X3D ran at 329 FPS AVG and slightly outperformed the AMD R5 7600 while roughly tying the Intel Ultra 5 245K. Intel manages better 1% lows than the 5500X3D. The 9800X3D holds a 500 FPS AVG result, a 53% improvement on the 5500X3D. Clearly there’s room to be better, but the last-gen 5700X3D and 5600X3D would be closer comparisons here. The 5600X3D outperformed the 5700X3D marginally from its frequency advantage, whereas the 5700X3D will be benefitted in more thread-intensive tasks. Both of them outdid the 5500X3D, with the 5600X3D improved by 11%. #### **Cyberpunk 2077 Phantom Liberty Benchmarks** Cyberpunk 2077: Phantom Liberty is up next. The 5500X3D ran at 164 FPS AVG, leaving the normal 5500 in the dust at 123 FPS AVG. The 5500X3D’s frametime pacing is also overall consistent and tied with the Intel 245K. The 265K is basically the same performance as the 5500X3D. AMD’s 5600X3D outperforms the 5500X3D by around 9%, the 5700X3D is almost 10% higher framerate, and the 5800X3D outdoes the 5500X3D by 17% for average framerate. The 9800X3D is 37% ahead, bouncing off of a GPU limit alongside the 9950X3D. The 5500X3D still outperforms the 14th generation and the non-X3D CPUs of multiple AMD generations, including the 9700X. X3D does well here and holds the entire top quarter of results. #### **Starfield Benchmarks** The 5500X3D does fine in Starfield as an objective measure, but isn’t competitive. The CPU runs at about 119 FPS AVG with unimpressive lows, although that’s mostly from the game. The CPU is at least significantly better than the R5 5500’s 88 FPS AVG and the 5600X’s 99 FPS AVG, despite allowing the 5600X3D a 10-11% lead. The 5700X3D pushes past the 5600X3D in this one, illustrating a benefit from the extra 2 cores even with the lower clock speed. The ceiling in this test is around 200 FPS AVG, with the 5800X3D being AM4’s closest CPU to that result at 148 FPS AVG, leading the 5500X3D by 24%. ### **5500X3D Production Benchmarks** #### **Blender** We’ll look at some production tests now. This won’t be as extensive as our Threadripper review, but will at least give us an idea as to the performance in non-gaming tasks. X3D doesn’t tend to help here, and in the AM4 generation, the configurations are often worse than non-X3D variants due to lower clocks. In Blender rendering of a 3D GN logo, the 5500X3D required about 30 minutes to complete the render of one frame from our intro animation. That has it slightly behind the 5500 non-X3D, although they’re about the same. The 5600X (watch our review) sees a reduction in time required of 9% thanks to its higher frequency. The 3700X outperforms all of these as a result of the 8-core configuration, which offers more in this workload than just frequency at lower core counts can. Overall, the 5500X3D performs like a low power 6-core CPU, which is what it is when the extra cache can’t be leveraged. The 12600K that it ranked alongside in several gaming tests is significantly and noticeably better here, up at 18 minutes for the same test. That’s a reduction in time required of 39%. #### **7-Zip Compression** In 7-Zip compression testing, the 5500X3D ran at about the same performance level as the 5600X3D. Results are in millions of instructions per second, with the 5500X3D completing 65K MIPS, the 5600X3D at about 70K MIPS, the regular 5500 at 64K MIPS, and the venerable 5800X3D completing 90K MIPS. The 5800X3D and 5700X3D both benefit from a higher core count, with the 5800X3D in particular at 38% ahead of the 5500X3D CPU. The 12600K is similar to the 5800X3D in this test. The 5500X3D isn’t impressive here in any capacity. There are better options for this kind of task, but if you had to do stuff like this in addition to budget gaming, it’d be workable. #### **7-Zip Decompression** Decompression is next. This test has the 5500X3D at 80K MIPS, landing just below the R5 5500, which benefits from the higher frequency that we showed earlier. The 12600K leads the 5500X3D by 18%, so where they were similar in a few games, the 12600K definitely has an advantage in these more core-intensive tasks. The 5500X3D is not a good CPU for this kind of work. It can get it done, but it’s not competitive. This is primarily a budget, cut-down gaming CPU. #### **Chromium** Chromium code compile is next. This test is extremely core-intensive and beats up the 5500X3D. It required 337 minutes to complete the code compile. That’s faster than the 355 minutes of the 5500 non-X3D, but still far down the chart. The 5600X and 5600X3D show similar rankings, with the 5700X3D also benefitted over the 5700X non-3D. There’s at least a clear trend in these AM4 CPUs of the extra cache helping versus their non-X3D counterparts. #### **Adobe Photoshop** In Adobe Photoshop with the Puget suite, the 5500X3D lands again at the bottom of the chart. It’s ahead of the 5500 non-3D but behind almost everything else, including the gaming neighbor 12600K. #### **Adobe Premiere** Adobe Premiere video editing and rendering tasks via the Puget suite have the 5500X3D at 6760 points extended, which plants it between the 5500 and 5600X again. The 3700X isn’t far off from the 5500X3D’s performance, mostly thanks to its extra 2 cores. The 12600K is significantly better than the 5500X3D here, as are basically any options from the AMD 7000 series and beyond. The generational uplift was particularly large in this test. #### **DaVinci Resolve** In DaVinci Resolve, which is experimental for us right now, the ranking is similar: The Puget suite keeps it down toward the bottom of the chart, just behind the 3700X, 5600X, and ahead of the 5500 non-X3D. ### **Conclusion** Grab a GN15 Large Anti-Static Modmat to celebrate our 15th Anniversary and for a high-quality PC building work surface. The Modmat features useful PC building diagrams and is anti-static conductive. Purchases directly fund our work! (or consider a direct donation or a Patreon contribution!) The CPU market is still relatively sane compared to the utter chaos that is the GPU market right now. CPUs are, for the most part, pretty stable—pricing is predictable, availability is decent, and while you might not agree with the price of a specific model, at least you know what you’re getting into. One of the most impressive things here is AM4. It’s still kicking, even with AM5 taking over. We’ve talked to motherboard vendors, and they’ve told us that older boards from past generations often get rerouted to markets outside the U.S., like South America and parts of Asia, where there is still demand. This helps keep prices a little lower and ensures that those boards don’t just end up in a landfill. And for folks looking to buy in, it means that entry costs for things like the 5500X3D are more reasonable. Speaking of the 5500X3D, it’s a noticeable improvement over the non-X3D 5500. The biggest problem with that old chip was its tiny cache. Performance-wise, the 5500X3D is fine overall and didn’t really fail in any of the games we tested. It isn’t a particularly good performer in things like production applications, compression, and decompression. It can do them, but it's not good if you’re doing those tasks heavily. In that case, you’d be better off going for a higher-core-count CPU. Pricing is key here, and it really depends on where you are. In the U.S., the 5500X3D’s launch price of $240 to $250 felt a little too steep, especially with better alternatives on the market at similar prices. But if you catch it around $200, it’s a much better deal and competes with stuff like the 9600X, so it’s definitely a more appealing option. What’s really impressive, though, is how AM4 is still relevant and kicking, despite being several years old. This thing has been around for ages, and AMD is still giving it support. AM4’s long life means less waste, fewer upgrades, and a platform that’s not forcing you to buy a whole new setup every couple of years. * * *

AMD Threadripper 9980X 64-Core CPU Review & Benchmarksjimmy_thang August 25, 2025 ## We compare the 9980X vs previous Threadripper CPUs (like the 7980X) and against desktop-class CPUs (like the Intel 285K and AMD Ryzen 9950X) The Highlights * AMD’s new 9980X Threadripper moves to the Zen5 architecture * On the gaming side, there are sometimes issues with consistency * The 9980X is anywhere from 2% to 58% improved upon the 7980X in our benchmarks * Original MSRP: $5,000 * Release Date: July 31, 2025 #### Table of Contents * AutoTOC Visit our Patreon page to contribute a few dollars toward this website's operation (or consider a direct donation or buying something from our GN Store!) Additionally, when you purchase through links to retailers on our site, we may earn a small affiliate commission. ### Intro We’re reviewing the $5,000 AMD Threadripper 9980X 64-core CPU and have a ton of new production tests to benchmark it. A lot of this is just an excuse to do some cool new testing. Like with medical simulations where we saw an 18% generational uplift, similar to the 18% we saw in financial options and black-scholes modeling, or the 58% improvement over the 7980X that we saw in convolution benchmarking. In most places, the gain is much smaller, like in compression with a couple percentage points of change. However, from the Zen 5 change to how AVX instructions are handled, it sometimes gains disproportionately. We saw the same with the 9900X vs. 7900X (or X3D variants) when in AVX-heavy scenarios. _Editor's note: This was originally published on July 30, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication._ * * * #### Credits * * * ##### Test Lead, Host, Writing Steve Burke ##### Testing Patrick Lathan Mike Gaglione ##### Video Editing Tim Phetdara ##### Writing, Web Editing Jimmy Thang * * * The quick version up front is that the improvement on the 7980X is anywhere from around 2% to that 58% number, but the vast majority of tests are closer to the range of 5% to 18%. It really heavily depends on the workload with these CPUs, particularly AVX occupancy. These are workstation CPUs intended for research, production work, or generally tasks that either make money or advance science or something similar. We’re still benchmarking gaming, but the focus will be mostly on production workloads today with gaming only there to ensure there are no major broken results. These CPUs are expensive. That limits the audience. For that reason, we added all the new tests as an opportunity to also re-run everything else, so all of this data is brand new. Even if you’re not in the market for Threadripper, this will have interesting data from more “normal” CPUs out there. ### **Overview** Here’s the quick recap of the CPU lineup. AMD is launching two versions of the Threadripper 9000 CPUs. These include the non-Pro and Pro CPUs. The Pro CPUs had an earlier embargo; today, we’re covering the HEDT (or “high-end desktop”) non-Pro Threadripper CPUs. The 9980X is a 64-core, 128-thread part that boosts up to 5.4 GHz and has 256MB of L3 Cache. The CPU sockets into STR5 platforms and, like the other two modern Threadripper parts launching with it, has a 350W TDP. It’s $5,000 for the 9980X. The 9970X is a 32-core, 64-thread part. They all share the same boost of 5.4GHz, but the base clock is higher on the lower core-count CPUs as a result of higher power availability per core based on power budget. Cache is down to 128MB L3 on the 9970X and 9960X, the latter of which is a 24-core, 48-thread part. Pricing is $2,500 for the 9970X and $1,500 for the 9960X. Compared to last time, prices are familiar: The 7980X was $5,000, the 7970X was $2,500, and the 7960X was $1,500. Threadripper 9000 prices are elevated from several generations back to the 3000 series, but the same as the 7000 series. The architectural improvements are the same as in Zen 5 with the 9000 series launch. The biggest one you’ll see today is from the AVX changes. When we were digging into the largest improvement in our testing, AMD sent this explanation as a refresher on Zen 5: “At first glance [the test] does appear to benefit from AVX-512 since the code has multiple references to ZMM. The main difference in 7K and 9K is the double vs single pump (256-bit native vs 512 native).” “High level, it took two cycles to do a 512 bit operation on Zen4, but one cycle on Zen5 because we doubled the FP datapath width, and increased the bandwidth from the L2 to the core to be able to match that speedup. You can't always take advantage of it, but theoretically there is a 2X improvement.” That “theoretically” is important, because it doesn’t show up in most of our tests; however, in some, there is a very large uplift. This isn’t special to Threadripper and is instead specific to Zen 5 vs. Zen 4. We’ll spend some time on that in our review. As for the test bench: Our chart subtitles apply to the desktop platforms, but the Threadripper CPUs are tested with different memory and motherboards. It deviates from some of the other benches. Let’s get into the benchmarks. We’ll start with production testing. ### 9980X Production Benchmarks #### **7-Zip Compression** 7-Zip compression is up now. This test is measured in MIPS, or millions of instructions per second. Higher is better throughput. The 9980X is the new chart-topper here, at 520K MIPS versus the 7980X’s 508K MIPS. The improvement is just 2.3% here. This is one of the lower improvements. There’s clear benefit to higher core counts in this test. The 7980X over the 7970X previously held a lead of 34%, or 67% over the 7960X. Against the highest-end desktop part, the 9950X3D, the 9980X sees an improvement of 151%. Notably, extra cache seems to help here: Both the 7950X3D and 9950X3D are benefitted over the 9950X. Threadripper is in an entirely different class from desktop components, although the uplift over the last generation’s 7980X is limited and relatively boring here. #### **7-Zip Decompression** In decompression, the 9980X scaled up to 926K MIPS, gaining 8% on the 7980X’s 857K MIPS. This is a better showing than what we saw in compression and posts better scaling. The rest of the lineup is familiar to the prior examples. Extra cache doesn’t help as much in this test as compared to the compression test, as the 7950X3D is now below the 9950X. This would align with the speed benefit providing more of an uplift on Threadripper as well. #### **SpecWS: OptionsPricing** SpecWorkstation 4.0 is next. This runs a number of simulations and mathematical models that we aren’t experts in, but we can at least run the benchmarks. We’ll rely on the Spec website to explain what the tests are. The first is Options Pricing. The website says that this runs Monte Carlo probabilistic simulations for financial uncertainty, Black-Scholes pricing models for theoretical value, and binomial options pricing. We are not experts in how these are used, but those of you in our audience who are experts have repeatedly expressed an appreciation for this test. The results have the 9980X at the top, with a 7.8 score. That has it improved on the 7980X by an impressive 18%, aligning with the findings of the 9000 series desktop parts against the 7000 series in this same test. You can see that in the 9800X3D against the 7800X3D or the 9900X3D vs. 7900X3D. This is an architectural benefit. It’s one of the rare break-outs for Zen 5 that we talked about after the initial round of reviews when we had validated it as a legitimate deviation. The 9980X improves on the highest-end desktop 9950X3D by 200%, basically being the difference between being runnable and not. Threadripper really makes an impact here, and the generational gain is real but rare in this one test. #### **SpecWS: OpenFOAM** OpenFOAM is next. Per the Spec website, OpenFOAM is a CFD test that “performs 2D Reynolds averaged simulation for fluid dynamics analysis.” It also runs a solver test. The 9980X scores 23.3 here, leading the 7980X prior generation by about 9%. Not as impressive as Options Pricing, but better than some other tests we have today. The lead over desktop parts like the 9950X and X3D is enough that it’s really not in the same class, with multiples of uplift. #### **SpecWS: Convolution** Convolution testing is next. Spec says that this test applies “a convolution filter to an image, a critical operation in signal processing and image analysis” for edge detection and feature extraction. This is heavily multi-threaded. In this test, the 9980X scored 9.3 to the 7980X’s 5.9, a huge uplift of 58%. This is completely out of alignment with any other tests. We contacted Wendell of Level1 Techs and AMD’s engineers to get both a first-party and a neutral third-party check of these numbers. Wendell noted that this is expected behavior and had seen similar results in other tests, believing it to be related to AVX improvements in addition to being a test that runs long enough to benefit from them. AMD noted that it has also seen results around 60% for certain similar tests. Although we have our 9970X review coming up separately, we’ll note that it also saw a 56% improvement here over the 7970X predecessor. The results align. Looking at last gen, the 9950X improved on the 7950X by 50%, which is consistent generationally. The results appear to be consistent and are present across numerous CPUs cross-generation, which means the only thing left to consider is whether this benchmark represents real-world scenarios. We’re not sure: This particular test exits our expertise, as we don’t do signal processing or image analysis work. We’d love to hear from those in our audience as to whether the Spec Convolution test aligns with real-world applications that you use in your work life. Regardless, the result is repeatable -- it’s only a question of whether it’s effectively a microbenchmark or is something that scales to real-world use. #### **SpecWS: LAMMPS** * * LAMMPS testing is next. According to Spec, this test is for large-scale atomic and molecular massively parallel simulation, which Spec says “utilizes MPI to scale to multiple cores to perform complex simulations.” This includes life science subtests, such as for polymer chain calculations and protein calculations for, what we assume are, medical uses. LAMMPS has the 9980X at a 5.5 score in Spec, leading the 7980X by 17%. This is close to the Options Pricing results. Although Spec has several tests where there’s no real impact, like Handbrake, it also has a higher saturation of tests with larger generational improvements. #### **SpecWS: Data Science** The Spec Data Science test is next. Spec says that this test “involves running a series of synthetic benchmarks that simulate real-world AI and ML workflows using the Python-based data science libraries Pandas, Scikit-learn, and XGBoost.” Generationally, the 9980X improves on the 7980X by 10%, going from 2.0 to 2.2 in the scoring. For reference, the 9950X improved on the 7950X by 13% at 1.7 from 1.5. #### **SpecWS: NAMD** NAMD is next, another test our audience liked before. This is a molecular dynamics simulator for modeling “behavior of biomolecular systems,” with Spec saying that it is “used for research in biochemistry, pharmacology, and molecular biology.” The 9980X scored 6.0, with the 7980X at 5.1. This is an improvement of almost 18%, which aligns again with LAMMPS and options pricing benchmarks. #### **SpecWS: RodiniaCFD** We have a lot more Spec results, but we’ll close on one that’s less interesting to bring it back down to earth and balance the higher results. This is for RodiniaCFD, which the site explains has a pre-euler subtest that “executes an unstructured grid finite volume CFD solver for 3D euler equations for compressible flow.” The 9980X scored 3.7 here, with the 7980X at 3.5. This is one where the improvement is less exciting, back down to about a 5-6% improvement generationally. #### **Chromium** Up next is Chromium code compile. There are a lot of types of code compile, just like there are a lot of types of video encoding or photo editing or 3D rendering workloads, so we’re just representing one here as we do with any other production test. The 9980X completes the compile the fastest, at 48 minutes total. The compile time is 7.5% reduced from the 7980X. Elsewhere, with the newer version of this test, we’re seeing reduced impact compared to prior test iterations. The 7980X and 7970X are relatively close to each other, with the 9950X3D and 9950X also near each other. Some of these CPUs are running with reduced memory speed as they required higher memory capacity and had more trouble holding our usual settings, so those are noted where true. The 14900K is Intel’s best performer, requiring 100 minutes to complete the compile. The 9950X outperforms this. #### **Blender** Blender tile-based rendering is up next. For this process, the 9980X is the new chart-topper at 2.1 minutes to complete a render of a single frame from the GN intro animation logo. The next fastest is the 7980X at 2.5 minutes, which makes sense: This test is heavily thread-dependent, so threads will outrank frequency where large differences emerge. Intel’s newer generation does OK in this test, with the 285K between the 9950X and 7950X. It’s also one of the tests where Intel doesn’t see regressive performance, though we still wouldn’t recommend the 285K. That entire platform is dead, anyway. #### **Photoshop** Adobe Photoshop is next, tested with the Puget Suite. This uses an aggregate score of filters, resizes, and other functions within Photoshop. The best performer here is the 9950X, followed by other desktop CPUs. The Threadripper parts fall further down the chart, with the 9980X in particular underperforming. We’ve seen this in the past, such as with the 7980X: They just have too many threads for their own good and Photoshop isn’t able to utilize them. The 7980X is outperformed by the 7970X and 7960X, so the 9980X landing in the middle isn’t a surprise. Its main improvement comes from the frequency. It does OK here if this is just one of many applications you use, but if Photoshop is the only thing you do, there are better and cheaper CPUs for the job. #### **Premiere** Adobe Premiere is next, our best frenemy that we see every day. In this one, the 9980X improves upon the performance from the 7980X and 7970X, but only barely. It’s definitely not worth a single generational upgrade, though we generally don’t recommend that anyway. The improvement is almost 5% from the 7980X to the 9980X in the extended score for Premiere. Intel’s 285K trails most Threadripper CPUs in the chart and roughly ties the 7960X. It also technically outperforms the 9950X. Intel’s biggest advantage remains in its QuickSync solution with the IGP, which would show up disproportionately in more targeted tests rather than the full sweep done here. Threadripper does well overall here and is at least better than its Photoshop showing, but without other applications leaning on threads, a desktop CPU would get most of the same performance. The best use for Threadripper here would be assigning its excess threads to other tasks, like separate audio compositing, while running Premiere tasks. #### **DaVinci Resolve** DaVinci Resolve is next. We haven’t tested this in years, so this is an experimental chart. In this test, the 9980X’s frequency boost allows it to gain on prior Threadripper CPUs. The previous lineup had the 7960X and 7980X tied, with the 7970X marginally ahead. We saw this with the 7000-series Threadripper reviews where 64-core parts sometimes caused slight losses compared to 32-core counterparts. The 285K falls behind the 9950X in this one, though not by a lot, with the 14900K roughly within re-run distance of the 285K. Threadripper is at least the best performer here as well, despite not providing clear value over desktop parts. It’s just not built for these kinds of workloads, but still being the best is important. ### **Gaming Benchmarks** Buy a GN 4-Pack of PC-themed 3D Coasters! These high-quality, durable, flexible coasters ship in a pack of 4, each with a fully custom design made by GN's team. You'll get a motherboard-themed coaster with debug display & reset buttons, a SATA SSD with to-scale connectors, RAM sticks, and a GN logo. These fund our web work! Buy here. We’ll keep gaming short. The only purpose of gaming tests on Threadripper, generally, is to make sure there isn’t some major problem, but cheaper desktop parts are going to be better for pure gaming builds. Threadripper is actually detrimental if all you do is gaming. You're going to notice that the Threadripper chips don't always follow a strictly logical order, especially when they're within a few frames per second of each other. That's because of the increased run-to-run variance causing less meaningful averaged results. #### **Stellaris Simulation Time** Stellaris simulation time is first. The 9800X3D is the current leader at 37.1 seconds on average for simulation time. The 9700X follows this at 43.2 seconds, benefitting from Zen 5 architectural improvements over the prior generation. The 285K isn’t what we’d call “good” here, but at least matches its predecessor. Threadripper is way at the bottom, with the 9980X at least improving on the 7980X notably, but it’s still far worse than any other modern gaming CPU. The 5700X is nearly at the level of the 9980X. It’s just detrimental to have this CPU configuration for this game, which isn’t really news to anyone. It can at least play games, but the difference in simulation time between the 9800X3D and the 9980X would be noticeable in real game play. #### **FFXIV: Dawntrail - 1080p** Final Fantasy 14: Dawntrail is next. The 9980X and 7980X high core-count CPUs had some major inconsistency issues in this game, so there’s an asterisk next to their entries. We’re still showing the averages of 4 runs, it’s just that the runs are more variable than typically. The 9980X, for instance, had a range from 274 to 248 FPS AVG, which is a huge swing. Our typical standard deviation in this game is a couple frames per second for the average. The game was playable, but all over the place for the average framerate. Lows were worse overall as well, but again, technically playable. This is why we still test games with Threadripper, though. #### **Dragon’s Dogma 2 - 1080p** In Dragon’s Dogma 2 at 1080p, the Threadripper CPUs end up in the lower half of the chart. The 9980X is worse than the 7950X here and about the same as the 7970X. The 7980X was below the 7960X and 7970X, indicating that these extremely high core counts are detrimental to gaming performance. The higher frequency matters more than the extra cores, so the trade-offs are harmful to gaming performance here. This has been true forever with Threadripper, so none of this is news. What matters more is that they’re still playable. Threadripper does suffer from worse lows, particularly 0.1%, but not so bad that they render the game stuttery. You could still play games on a Threadripper system, but again, gaming-only users should buy something else. #### **Cyberpunk 2077: Phantom Liberty - 1080p** Cyberpunk is next. In this one, the 9980X outperforms the 7980X on a technicality, but both are closer to the R7 2700 than they are to the 9800X3D. The run-to-run consistency was fine, with the 9980X only exhibiting a 1.9 AVG FPS standard deviation, so it’s not like what we saw in Final Fantasy -- it’s just not particularly good. This is still a passable framerate, so in the context of a computer built for specific work tasks that can still play games when needed, it passes that need. #### **Baldur’s Gate 3 - 1080p** In Baldur’s Gate 3, the 9980X ends up around the level of the 265K -- which is really just more embarrassing for Intel than anything else. We already know Threadripper isn’t particularly good at gaming, despite being OK at it, but the 265K should be doing better here. This is also a game where the 285K was hugely regressive, though, at 109 FPS AVG against the 14700K’s 117 FPS AVG and 14900K’s 123 FPS AVG. We did test other games, but they’re just not relevant beyond proving the point we already have: Threadripper is passable if gaming is not your primary need. It’s just not the best choice for it. ### Power Consumption #### **9980X Power Consumption - Blender** This next chart shows the power consumption as measured at the EPS12V cables with a power interposer between the motherboard and power supply. Tested with an all-core workload in Blender, we observed stable and consistent power draw across both the 7980X and 9980X CPUs, with the 9980X measuring at 371W from 354W on the last-gen Threadripper flagship. That’s a lot of power to cool. This does not show transients. ### 9980X Thermals #### **Thermals - Blender** This chart shows the thermal behavior of the 9980X. We’re using a SilverStone 360mm liquid cooler with a Threadripper-sized coldplate for this, so the results are not comparable to our 7980X review thermal results from its launch. Also, as usual, thermal testing would be best done with a lot of coolers compared and noise-normalized, but this is kept simple for the review and to help get an idea (as that’d be a cooler review). The 9980X under an all-core workload ran at around 58 degrees Celsius when running the relatively loud 100% fan speeds on the 360mm liquid cooler. #### **Thermals - Equilibrium** This chart shows thermals at steady state across the various CCDs. The CCD-to-CCD delta here is 6.7 degrees Celsius, at 49.7 to 56.1 degrees in an ambient of 21 degrees Celsius, +/- 1C. Generally speaking, this can be cooled at lower noise levels with 360mm coolers without major issues if assuming a well-ventilated case. The surface area is so huge on this CPU that the power is distributed across a large area, making it easier to cool than a 200W heat load might be on a normal desktop-sized Ryzen CPU with 2 CCDs. ### 9980X Frequency #### **Frequency - Blender All-Core** With all cores working, the CPU typically ran at around 3950 MHz, with occasional spikes to 4200 MHz while some threads were bringing in data for new tiles to render. The base clock is 3.2 GHz, so it’s not dropping that low during this particular workload, but it’s also far away from 5.4 GHz. This is typical behavior for all-core workloads. #### **Frequency - Blender vs 7980X** This chart shows the 7980X for the same test. The frequency bottoms-out at around 3.7 GHz, but is all over the place during this workload. The 9980X has improved upon the frequency in all-core scenarios overall. #### **Frequency - CB Single-Thread** In a single-threaded test with Cinebench, the 9980X ran at 5425 MHz, exceeding the spec listing of 5.4 GHz for boost. This is also expected behavior, with the more limited load boosting higher. The 7980X in the same test ran at about 5340 MHz, so the single-thread frequency has improved generationally -- but not much, and that’s why some benchmarks only show a couple percentage points of change. ### **9980X Conclusion** Grab a GN15 Large Anti-Static Modmat to celebrate our 15th Anniversary and for a high-quality PC building work surface. The Modmat features useful PC building diagrams and is anti-static conductive. Purchases directly fund our work! (or consider a direct donation or a Patreon contribution!) On the gaming side, there are sometimes issues with consistency. This means, run-to-run, you don’t get the same experience. There’s far more deviation here in average frame rate and the frametime consistency is worse than traditional desktop parts. This is especially true with 64-core Threadripper CPUs. This is not really unexpected. These high core count CPUs sometimes just have so many cores that it starts causing scheduling problems. Sometimes you’ll also end up in situations where, because the clocks aren’t boosting as high (because of the higher core count), you’re losing that frequency advantage. Again, this is not unexpected. The good news is that the games we tested on the 9980X weren’t broken. They were able to run and weren’t awful experiences, which wasn’t the case with the earliest Threadripper CPUs. Compared to the 32-core Threadripper, the 64-core part technically runs cooler. That’s because you’re taking the same power and distributing it across more silicon. This means you’re going to get lower temperatures per die per hot spot. That remains true here where the 64-core CPU can be cooler than the 32-core processor. With its prices, these Threadripper CPUs aren’t something you’re buying for “value.” The target demographic for these processors is probably making money with their computers. Threadripper also has the benefit of PCIe lane availability for machines with multiple accelerators or a lot of I/O, which are more limited with traditional desktop CPUs. 64 cores don’t always scale well. Sometimes a 32-core CPU can be a better fit. Make sure to do research online for your needs. For example, we found out that a 32-core CPU would work better for us while using Adobe Premiere for video production. So it’s best to research these CPUs with your needs. * * *

TIMELINE: GPU Export Controls, NVIDIA GPU Bans, & AI GPU Black Marketjimmy_thang August 18, 2025 ## We’ve compiled a comprehensive timeline of the GPU bans, GPU smuggling, and export controls that impact NVIDIA, AMD, and Intel The Highlights * The US blocks exports of advanced GPUs to China to protect national security * NVIDIA GPUs are highly sought after in China for AI processing * Our timeline chronicles the US export controls, NVIDIA's responses, and reports of GPU smuggling Visit our Patreon page to contribute a few dollars toward this website's operation (or consider a direct donation or buying something from our GN Store!) Additionally, when you purchase through links to retailers on our site, we may earn a small affiliate commission. This is a comprehensive timeline of the GPU bans, smuggling, and export controls on NVIDIA, AMD, Intel, and other high-tech semiconductor products. We are publishing this as part of our stretch goals for Black Market AI GPU — a viewer-funded film made possible through support on our store, including our new “Blind Eye” T-shirt. The below timeline accompanies our Black Market AI GPU investigation, our biggest project yet. We spent three weeks in Asia to uncover this story, including two weeks in China and one in Taiwan. We found smugglers, middlemen, and users of so-called “AI” GPUs that the United States government has banned for sale into China. * * * #### Credits * * * ##### Host, Writing, Lead Editing Steve Burke ##### Editing Vitalii Makhnovets Tim Phetdara ##### Editing, Graphics Andrew Coleman ##### Camera Tannen Williams ##### Research and Writing Ben Benson * * * We are providing this timeline for free and without third-party ads for our viewers and readers. As this situation has changed frequently and now spans multiple US administrations, we may have missed a few events. However, we believe we have compiled all the major changes – especially since the start of 2025 – that are directly relevant to the story. We have attempted to present it as neutrally as feasible and from a place of reporting. We’ve included links to a variety of media and government sources that we believe are appropriate for establishing the timeline of events. We have included statements from NVIDIA in many cases. This was a huge team effort at GN and required a massive investment in travel, writing, research, and editing to complete. If you find this information valuable, we ask that you please support us directly by backing our NVIDIA AI GPU Black Market project, buying something from our GN store, or signing up for our Patreon. Thank you. ### **Timeline** Note on sources: Our intent is to cite primary sources, including government documents, and a variety of secondary sources. In some cases, we link only to secondary news stories. This can occur when we include articles from credible media reports but do not have primary documents to cite. * * * #### 2018 GPU Export Controls ##### August 2018 August 13: The US government created the National Security Commission on Artificial Intelligence (NSCAI) as part of the John S. McCain National Defense Authorization Act for Fiscal Year 2019. The NSCAI had 15 commissioners who were nominated by Congress and the Executive Branch. The NSCAI was tasked with investigating how the United States should compete in AI in the modern age and recommending actions for Congress and the executive branch. In the words of the original document, the commissioners “shall consider the methods and means necessary to advance the development of artificial intelligence, machine learning, and associated technologies by the United States to comprehensively address the national security and defense needs of the United States.” * Source: * National Defense Bill (NSCAI section starts on page 1963) #### 2019 GPU Export Controls ##### May 2019 May 15: Citing national security risks, the US government added Huawei to its Entity List and restricted sales of Huawei’s equipment into the United States. * Sources: * Federal Register * Executive Order * Politico: Trump signs order setting stage to ban Huawei from U.S. * Reuters: Trump administration hits China's Huawei with one-two punch * Reuters: China's Huawei, 70 affiliates placed on U.S. trade blacklist #### 2020 GPU Export Controls ##### May 2020 May 19: The United States restricted semiconductor designs, chipsets, and technologies to Huawei and its foreign affiliates. * Sources: * Department of Commerce Press Release * Federal Register * NPR: The Latest U.S. Blow To China's Huawei Could Knock Out Its Global 5G Plans #### 2021 GPU Export Controls ##### March 2021 The National Security Commission on Artificial Intelligence (NSCAI) released its final report. The report provided recommendations to “advance the development of artificial intelligence, machine learning, and associated technologies to comprehensively address the national security and defense needs of the United States.” As part of the report (page 216), the NSCAI recommended the US government and its allies “utilize targeted export controls on high-end semiconductor manufacturing equipment… to protect existing technical advantages and slow the advancement of China’s semiconductor industry.” Further on (page 228), the report said, “Looking across the AI stack, the hardware component of the AI stack contains the most viable targets for traditional export controls.” The report (page 231) focused on semiconductor manufacturing equipment for export control rules: “The primary U.S. export control target to constrain competitors’ AI capabilities should be sophisticated semiconductor manufacturing equipment (SME) necessary to manufacture high-end chips.” The report mentioned export controls for GPUs (page 500) as a way “to prevent the use of high-end U.S. AI chips in human rights violations.” * Source: * NSCAI Final Report – The PDF can be downloaded from here. ##### April 2021 NSCAI Commissioner Christopher Darby spoke at NVIDIA GTC about the NSCAI’s report to Congress. * Source: * GTC ##### October 2021 October 1: The NSCAI officially ended on October 1, 2021. * Source: * Executive Government News #### 2022 GPU Export Controls ##### August 2022 August 31: NVIDIA filed a Form 8-K with the SEC to inform investors that the US government had immediately blocked exports of its A100 and H100 chips to China, including Hong Kong. The export controls included DGX or other systems that incorporate an A100, H100, or A100X. In the financial documents, NVIDIA said the US government informed it of the export restrictions on August 26, 2022. NVIDIA stated that its third-quarter results included up to $400 million in expected sales to China that were now uncertain due to the export restrictions. * Sources: * NVIDIA Form 8-K * NVIDIA Form 10-Q * SEC Filing ##### September 2022 September 1: NVIDIA filed a new Form 8-K to let customers know that the US government had offered some exemptions for certain chip exports: “The U.S. government has authorized exports, reexports, and in-country transfers needed to continue NVIDIA Corporation’s, or the Company’s, development of H100 integrated circuits after the Company filed its Current Report on Form 8-K with the U.S. Securities and Exchange Commission on August 31, 2022. The authorization also allows the Company to perform exports needed to provide support for U.S. customers of A100 through March 1, 2023. Additionally, the U.S. government authorized A100 and H100 order fulfillment and logistics through the Company’s Hong Kong facility through September 1, 2023.” * Source: * NVIDIA Form 8-K Following NVIDIA’s SEC filing, media outlets reported the US government ordered NVIDIA to stop selling advanced AI chips to China. * NVIDIA statement: * “We are working with our customers in China to satisfy their planned or future purchases with alternative products and may seek licenses where replacements aren’t sufficient. The only current products that the new licensing requirement applies to are A100, H100 and systems such as DGX that include them.”– NVIDIA to CNBC * Sources: * CNBC: NVIDIA stock falls after U.S. government restricts chip sales to China * The New York Times: U.S. Restricts Sales of Sophisticated Chips to China and Russia * Reuters: U.S. officials order NVIDIA to halt sales of top AI chips to China * CNN: US orders NVIDIA and AMD to stop selling AI chips to China * Associated Press: China demands US drop tech export curbs after NVIDIA warning ##### October 2022 October 7: The Department of Commerce’s Bureau of Industry and Security (BIS) implemented a series of export controls to “protect US national security and foreign policy interests.” The new export controls would hinder China’s ability to build high-end semiconductors and purchase advanced chips from the US, including for development of and maintaining supercomputers. * Sources: * Department of Commerce Release * Federal Register (amended on October 13) In a briefing with reporters, the US government said the new regulations formalized the guidance previously sent to NVIDIA. The Guardian reported: “The new regulations will also severely restrict export of US equipment to Chinese memory chip makers and formalize letters sent to NVIDIA Corp and Advanced Micro Devices Inc (AMD) restricting shipments to China of chips used in supercomputing systems that nations around the world rely on to develop nuclear weapons and other military technologies.” * Sources: * The Guardian: Biden administration imposes sweeping tech restrictions on China * The New York Times: Biden Administration Clamps Down on China’s Access to Chip Technology ##### November 2022 November 7: Reuters reported that NVIDIA had created a new AI chip called the A800 GPU for the China market. The A800 would be compliant with US export controls. * NVIDIA statement: * “The NVIDIA A800 GPU, which went into production in Q3, is another alternative product to the NVIDIA A100 GPU for customers in China. The A800 meets the US government’s clear test for reduced export control and cannot be programmed to exceed it.” - NVIDIA to Reuters * Sources: * Reuters: Exclusive: NVIDIA offers new advanced chip for China that meets U.S. export controls * Tom’s Hardware: NVIDIA Creates New Supercomputer Chip For Chinese Market * The Verge: NVIDIA’s selling a nerfed GPU in China to get around export restrictions * TechCrunch: NVIDIA touts a slower chip for China to avoid US ban #### 2023 GPU Export Controls ##### March 2023 March 21: Reuters reported that NVIDIA had modified the H100 to be compliant with export rules to China. * NVIDIA’s statements to Reuters: * “On Tuesday, the company said it has similarly developed a China-export version of its H100 chip. The new chip, called the H800, is being used by the cloud computing units of Chinese technology firms such as Alibaba Group Holding Ltd, Baidu Inc and Tencent Holdings Ltd, a company spokesperson said.” [...] “The NVIDIA spokesperson declined to say how the China-focused H800 differs from the H100, except that ‘our 800 series products are fully compliant with export control regulations.’” * Sources: * Reuters: NVIDIA tweaks flagship H100 chip for export to China as H800 * Data Center Dynamics: NVIDIA creates pared back H100 GPU for export to China, called H800 ##### June 2023 June 27: The Wall Street Journal reported that the US government is considering expanding export controls for GPUs and AI chips to China. The US Department of Commerce did not comment to the Wall Street Journal. * Sources: * The Wall Street Journal: U.S. Considers New Curbs on AI Chip Exports to China * MarketWatch: White House says it’s focused on being at front end of supply chain for chips, won’t comment on report of possible new ban on exporting AI chips to China * Reuters: US mulls new export restriction on computing power in AI chips * TechCrunch: China’s AI firms might further lose chip access in new US ban ##### October 2023 October 17: The US Department of Commerce updated its export compliance for advanced semiconductors and semiconductor manufacturing equipment. The government said: “Today’s rules reinforce the October 7, 2022, controls to restrict the PRC’s ability to both purchase and manufacture certain high-end chips critical for military advantage. These updates are necessary to maintain the effectiveness of these controls, close loopholes, and ensure they remain durable.” The U.S. government has removed interconnect speed as a criterion for identifying restricted chips. Instead, it will now focus on processor performance and performance density. In a statement, the government said: “A performance density parameter prevents the workaround of simply purchasing a larger number of smaller datacenter AI chips which, if combined, would be equally powerful as restricted chips.” * Sources: * Department of Commerce Release * Federal Register * Bureau of Industry and Security Document on Performance Density (page  21-22) * The Register: Biden has brought the ban hammer down on US export of AI chips to China * Center for Strategic & Internal Studies (posted on Oct. 18) As part of the announcement, the administration told reporters the new restrictions affect NVIDIA’s A800 and H800 chips. A few days prior, Reuters reported that the administration would soon announce new export rules. * Sources: * Reuters: Exclusive: US tackles loopholes in curbs on AI chip exports to China * CNBC: U.S. curbs export of more AI chips, including NVIDIA H800, to China * Reuters: NVIDIA details advanced AI chips blocked by new export controls * Tom’s Hardware: US Prohibits Exports of NVIDIA’s A800 and H800 to China, Blacklists Chinese GPU Developers * The Verge: NVIDIA’s H800 AI chip for China is blocked by new export rules October 23: NVIDIA filed a Form 8-K with the SEC that said the new export rules impact its A100, A800, H100, H800 and L40S chips. NVIDIA said it “does not anticipate that the accelerated timing of the licensing requirements will have a near-term meaningful impact on its financial results.” * Sources: * NVIDIA Form 8-K * BBC: US orders immediate halt to some AI chip exports to China, NVIDIA says * Tom’s Hardware: US Govt Speeds Up Export Restrictions for NVIDIA’s GPUs ##### December 2023 Grab a GN15 Large Anti-Static Modmat to celebrate our 15th Anniversary and for a high-quality PC building work surface. The Modmat features useful PC building diagrams and is anti-static conductive. Purchases directly fund our work! (or consider a direct donation or a Patreon contribution!) December 6: In a meeting with reporters in Singapore, NVIDIA said that it was working on new chips that comply with the government’s rules. * NVIDIA statement: * “NVIDIA has been working very closely with the U.S. government to create products that comply with its regulations. Our plan now is to continue to work with the government to come up with a new set of products that comply with the new regulations that have certain limits.” – NVIDIA CEO Jensen Huang, as reported in Reuters * Sources: * Reuters: NVIDIA working closely with US to ensure new chips for China are compliant with curbs * Reposted in CNBC December 28: NVIDIA released a new version of RTX 4090 for the China market. The new chip, called the GeForce RTX 4090D, would be compliant with US export control restrictions. * NVIDIA statements: * “The GeForce RTX 4090 D has been designed to fully comply with U.S. government export controls. While developing this product, we extensively engaged with the U.S. government.” - NVIDIA to Reuters * “In 4K gaming with ray tracing and deep-learning super sampling (DLSS), the GeForce RTX 4090D is about five percent slower than the GeForce RTX 4090 and it operates like every other GeForce GPU, which can be overclocked by end users.” – NVIDIA to The Register * Sources: * Reuters: NVIDIA launches new gaming chip for China to comply with US export controls * The Register: NVIDIA slowed RTX 4090 GPU by 11 percent, to make it 100 percent legal for export to China * The Verge: NVIDIA is releasing a slower RTX 4090 in China to comply with US restrictions #### 2024 GPU Export Controls ##### February 2024 February 1: Reuters reported that NVIDIA had prepared new GPUs for China, including the H20. Several sources told Reuters that the new offerings are less powerful than similar chips from Huawei. * Sources: * Reuters: Exclusive: NVIDIA’s new China-focused AI chip set to be sold at similar price to Huawei product * Tom’s Hardware: New NVIDIA AI GPUs designed to get around U.S. export bans come to China — H20, L20, and L2 to fill void left by restricted models ##### July 2024 July 22: Reuters reported that NVIDIA is creating a new GPU for the China market based on its Blackwell chips. Sources told Reuters that the chip would be a version of the Blackwell B200. NVIDIA did not publicly disclose the specifications. * Sources: * Reuters: Exclusive: NVIDIA preparing version of new flagship AI chip for Chinese market * Tom’s Hardware: NVIDIA preparing a China-focused variant of its B200 Blackwell AI GPU to comply with US export regulations * HPCWire: NVIDIA Prepares New AI Chip for China Amid Ongoing US Export Controls ##### December 2024 December 2: The US government expanded rules that limit the export of high memory bandwidth (HBM) and advanced semiconductor equipment. The Center for Strategic & International Studies explained the new rules on HBM: “The December 2024 controls change that by adopting for the first time country-wide restrictions on the export of advanced HBM to China as well as an end-use and end-user controls on the sale of even less advanced versions of HBM. The goal of these controls is, unsurprisingly, to degrade China’s AI industry.” [...] “Modern AI chips not only require a lot of memory capacity but also an extraordinary amount of memory bandwidth. Bandwidth refers to the amount of data a computer’s memory can transfer to the processor (or other components) in a given amount of time. With low-bandwidth memory, the processing power of the AI chip often sits around doing nothing while it waits for the necessary data to be retrieved from (or stored in) memory and brought to the processor’s computing resources.” * Sources: * Department of Commerce Document * Federal Register * Government Presentation * Center for Strategic & International Studies: Understanding the Biden Administration’s Updated Export Controls #### 2025 GPU Export Controls ##### January 2025 January 13: The US government tightened its export controls by introducing national chip caps for many countries, except for 18 allies. The new restrictions would be called the AI Diffusion Rule. The rule would go into effect in May 2025. * NVIDIA statement: * “It makes no sense for the Biden White House to control everyday datacenter computers and technology that is already in gaming PCs worldwide, disguised as an anti-China move. The extreme ‘country cap’ policy will affect mainstream computers in countries around the world, doing nothing to promote national security but rather pushing the world to alternative technologies. AI is mainstream computing – ubiquitous and essential as electricity. This last-minute Biden Administration policy would be a legacy that will be criticized by U.S. industry and the global community. We would encourage President Biden to not preempt incoming President Trump by enacting a policy that will only harm the U.S. economy, set America back, and play into the hands of U.S. adversaries.” – Ned Finkle, Vice President of Government Affairs, NVIDIA, to Bloomberg (Twitter link) * Sources: * US Government Fact Sheet * Federal Register * Federal Register * Bloomberg: Biden to Further Limit NVIDIA AI Chip Exports in Final Push * Tom’s Hardware: NVIDIA and SIA fire back at US gov's new export restrictions on AI GPUs to China ##### February 2025 February 26: NVIDIA filed its 10-K annual report with the SEC. In the 10-K, NVIDIA revealed that Singapore was the second-largest geographical source of revenue in 2024, behind the United States. Taiwan was third, and China was fourth. Within the report, NVIDIA said: “Singapore represented 18% of fiscal year 2025 total revenue based upon customer billing location. Customers use Singapore to centralize invoicing while our products are almost always shipped elsewhere. Shipments to Singapore were less than 2% of fiscal year 2025 total revenue.” * Source: * NVIDIA 10-K February 27: Speculation began about AI GPUs being smuggled from Singapore to China. In late February, authorities in Singapore arrested three people for fraud involving servers that may contain AI GPUs. Singapore’s government granted the three people bail a few weeks later. NVIDIA declined to comment to CNBC. * Sources: * ChannelNewsAsia: 3 men charged with fraud, cases linked to alleged movement of Nvidia chips * CNBC: NVIDIA’s unofficial exports to China face scrutiny after arrest of silicon smugglers in Singapore * Tom’s Hardware: Singapore police bust major ring smuggling NVIDIA GPUs to China-based DeepSeek: Report * TechCrunch: Singapore grants bail for NVIDIA chip smugglers in alleged $390M fraud ##### April 2025 April 9: NPR reported that the US government would not add export controls for the H20 chip after NVIDIA CEO Jensen Huang attended a dinner at Mar-A-Lago. The dinner reportedly cost $1 million per head. The outlet said it was unclear whether Jensen Huang met with US President Trump directly. NVIDIA declined to comment to NPR. * Sources: * NPR: Trump administration backs off NVIDIA H20 chip crackdown after Mar-a-Lago dinner April 15: In a SEC filing, NVIDIA said the US government sent the company new export rules on April 9. According to NVIDIA, the H20 and all chips with the H20’s memory bandwidth or interconnect bandwidth will now need licenses to export to China. NVIDIA said the new rules would cost the company $5.5 billion in charges due to current H20 chip inventory and prior sales. NVIDIA declined to comment further to the BBC. * Sources: * NVIDIA Form 8-K * BBC: NVIDIA shares plunge amid $5.5bn hit over export rules to China * NPR: NVIDIA discloses that U.S. will limit sales of advanced chips to China after all * Reuters: US issues export licensing requirements for NVIDIA, AMD chips to China April 16: The US government released an investigative report on DeepSeek and requested information from NVIDIA about its AI GPUs. Through a letter sent to NVIDIA CEO Jensen Huang, the US government asked NVIDIA for a list of its customers in China and many countries in Asia, including Singapore. The government requested all communication between NVIDIA and DeepSeek. * Sources: * US government press release * DeepSeek report * Government letter to Jensen Huang The US Department of Commerce confirmed that it has issued new export control rules for AI chips. The Commerce Department provided a statement to The Wall Street Journal: “The Commerce Department is issuing new export licensing requirements on the NVIDIA H20, AMD MI308, and their equivalents.” * Sources: * The Wall Street Journal: U.S. Confirms New Export Curbs on NVIDIA and AMD Chips April 28: The Wall Street Journal reported that Huawei is expected to release its new AI chip, the Ascend 910D, soon. According to the Wall Street Journal’s sources, Huawei expects the Ascend 910D to be about as powerful as an NVIDIA H100. * Sources: * The Wall Street Journal: China’s Huawei Develops New AI Chip, Seeking to Match NVIDIA * NetworkWorld: Huawei steps up AI chip race with Ascend 910D, targeting NVIDIA’s high ground April 30: Anthropic, an AI startup backed by Amazon, called on the US government to increase export control restrictions to China. As part of a blog post, Anthropic said the government needs to improve its export enforcement to reduce smuggling. The company cited examples of chips being smuggled with “prosthetic baby bumps” and “live lobsters.” In a response, NVIDIA said: “American firms should focus on innovation and rise to the challenge, rather than tell tall tales that large, heavy, and sensitive electronics are somehow smuggled in ‘baby bumps’ or ‘alongside live lobsters.’” – NVIDIA to CNBC * Sources: * Anthropic blog post * CNBC: NVIDIA says Anthropic is telling ‘tall tales’ in its defense of U.S. AI chip restrictions on China * Tom’s Hardware: Despite NVIDIA claims, Chinese smugglers have used live lobsters and fake baby bumps to traffic chips * Hong Kong Customs release * China Customs release  ##### May 2025 May 1: Jensen Huang spoke with the House Foreign Affairs Committee to discuss domestic manufacturing and the importance of AI. NVIDIA posted the remarks online. * Sources: * NVIDIA Newsroom Twitter Post * Tom’s Hardware: NVIDIA warns U.S. AI hardware export rules could backfire, empowering Huawei to define global standards May 7: Following a report in Bloomberg, the US Department of Commerce confirmed that it will not implement the AI Diffusion Rule that was created during the prior administration. The rule would have gone into effect on May 15, 2025. The Department of Commerce released a statement to CNBC: “The Biden AI rule is overly complex, overly bureaucratic, and would stymie American innovation. We will be replacing it with a much simpler rule that unleashes American innovation and ensures American AI dominance.” NVIDIA released a statement: “We welcome the Administration’s leadership and new direction on AI policy. With the AI Diffusion Rule revoked, America will have a once-in-a-generation opportunity to lead the next industrial revolution and create high-paying U.S. jobs, build new U.S.-supplied infrastructure, and alleviate the trade deficit.” * Sources: * Bloomberg: Trump to Rescind Global Chip Curbs, Prep New AI Restrictions * Tom’s Hardware: NVIDIA celebrates dumping of Biden-era AI chip export rules — simpler new policy promised * CNBC: Trump administration set to end Biden’s U.S. chip export restrictions * NVIDIA Twitter Account May 9: Reuters reported that NVIDIA is preparing a cut down version of the H20 for the Chinese market. Reuters sources said the chip would be ready in July. NVIDIA declined to comment. * Sources: * Reuters: Exclusive: NVIDIA modifies H20 chip for China to overcome US export controls, sources say * Tom’s Hardware: NVIDIA readies cut-down HGX H20 GPU for China to comply with export control rules May 13: The US government formally rescinded the previous administration’s AI Diffusion Rule, which was announced in January 2025. The government also announced actions to strengthen export controls for AI chips, including restrictions on using several Huawei Ascend chips NVIDIA declined to comment to CNBC on the new export restrictions. * Sources: * US government press release  * BIS policy statement  * US government guidance on using Huawei Ascend chips * The Register: Trump ends Biden-era dream to cap US AI chip exports * Bloomberg: US Warns That Using Huawei AI Chip ‘Anywhere’ Breaks Its Rules * Tom’s Hardware: U.S. issues worldwide crackdown on using Huawei Ascend chips, says it violates export controls * CNBC: Trump administration’s next wave of China AI chip export rules are yet another obstacle for NVIDIA May 15: A bipartisan group of legislators introduced the Chip Security Act that is intended to stop smuggling of high-end AI chips. The Hill summarized the proposed legislation: “The legislation, titled the Chips Security Act, would require companies to ensure the location-verification abilities of their high-end AI chips and to report when a product has been diverted or changed location. It follows recent reports of increased smuggling of chips, including those made by NVIDIA, into China despite tight export controls.” NVIDIA declined to comment to The Register. * Sources: * Chip Security Act text * The Hill: Bipartisan House lawmakers propose bill to ‘stop smuggling’ of AI chips * Reuters: U.S. lawmakers introduce bill to address AI chip smuggling * The Register: Plan to keep advanced chips from China with tracking tech gains support in Congress May 16: The Financial Times reported that NVIDIA intends to create a research and design center in Shanghai. * NVIDIA statement: * “We are not sending any GPU designs to China to be modified to comply with export controls.” - NVIDIA to CNBC  * Sources: * Financial Times: NVIDIA plans Shanghai research centre in new commitment to China * CNBC: NVIDIA says it is not sending GPU designs to China after reports of new Shanghai operation May 19: NVIDIA CEO Jensen Huang told Bloomberg in a TV interview that he didn’t see any “evidence” of any AI chip diversion. Tom’s Hardware summarized Jensen Huang’s quote: “Governments understand that diversion is not allowed, and there’s no evidence of any AI chip diversion — recognize our data center GPUs are massive; these are massive systems. The Grace Blackwell system is nearly two tons, and so you’re not going to be shipping — you’re not going to be putting that in your pocket or your backpack anytime soon. And so, these systems are fairly easy to keep track of... but the important thing is that the countries and the companies that we sell to recognize that diversion is not allowed, and everybody would like to continue to buy NVIDIA technology, and so they very well monitor themselves very carefully and they’re quite careful about that.” * Sources: * Bloomberg: NVIDIA CEO Says ‘No Evidence of Any AI Chip Diversion’ * Tom’s Hardware: NVIDIA CEO Jensen Huang says ‘There’s no evidence of any AI chip diversion’ May 21: At Computex 2025, NVIDIA CEO Jensen Huang called the US export controls a “failure.” He said that NVIDIA’s market share in China has dropped from 95% to 50% due to the restrictions. The Guardian quoted Jensen Huang as saying: “The local companies are very, very talented and very determined, and the export control gave them the spirit, the energy and the government support to accelerate their development.” [...] “I think, all in all, the export control was a failure.” [...] “China has a vibrant technology ecosystem, and it’s very important to realise that China has 50% of the world’s AI researchers, and China is incredibly good at software.” * Sources: * CNBC: Jensen Huang says U.S. chip restrictions have cut NVIDIA’s China market share nearly in half * The Guardian: US chip export controls are a ‘failure’ because they spur Chinese development, NVIDIA boss says * The New York Times: NVIDIA’s Chief Says U.S. Chip Controls on China Have Backfired May 27: Reuters reported that NVIDIA plans to launch a new, cheaper Blackwell-based GPU for the China market to comply with US export rules. * NVIDIA statement: * "Until we settle on a new product design and receive approval from the U.S. government, we are effectively foreclosed from China's $50 billion data center market." - NVIDIA to Reuters * Sources: * Reuters: Exclusive: NVIDIA to launch cheaper Blackwell AI chip for China after US export curbs, sources say * SiliconANGLE: Report: NVIDIA racing to develop new, scaled-down Blackwell GPUs for China May 28: During NVIDIA’s quarterly earnings, CEO Jensen Huang said the company was writing off unsold H20 inventory due to export controls. VentureBeat posted Jensen Huang’s quote from earnings: “Let me share my perspective on some topics we’re frequently asked on export control. China is one of the world’s largest AI markets and a springboard to global success with half of the world’s AI researchers based there. The platform that wins China is positioned to lead globally today. However, the $50 billion China market is effectively closed to U.S. industry. The H20 export ban ended our Hopper data center business in China. We cannot produce Hopper further to comply. As a result, we are taking a multibillion-dollar write-off on inventory that cannot be sold or repurposed. We are exploring limited ways to compete, but hopper is no longer an option.” * Sources: * VentureBeat: NVIDIA CEO takes a shot at U.S. policy cutting off AI chip sales to China * PC Gamer: NVIDIA’s Hopper GPUs are now dead to the Chinese market after export controls that made the company take a 'multibillion-dollar write-off' ##### June 2025 Our fully custom 3D Emblem Glasses celebrate our 15th Anniversary! We hand-assemble these on the East Coast in the US with a metal badge, strong adhesive, and high-quality pint glass. They pair excellently with our 3D 'Debug' Drink Coasters. Purchases keep us **ad-free** and directly support our consumer-focused reviews! June 12: NVIDIA’s CEO Jensen Huang told CNN the company will no longer include sales and revenue from China in its forecasts. In a response to a question from CNN about whether the US government would lift its export controls, Jensen Huang said: “I’m not counting on it but, if it happens, then it will be a great bonus. I’ve told all of our investors and shareholders that, going forward, our forecasts will not include the China market.” * Sources: * CNN: NVIDIA will stop including China in its forecasts amid US chip export controls, CEO says June 18: Several media reported on rumors about NVIDIA preparing to launch a “RTX 5090 DD” for the China market. The new card would allegedly reduce the memory specifications compared to the RTX 5090D. * Sources: * Tom’s Hardware: NVIDIA planning new RTX 5090 'DD' variant for China — 24GB card with tweaked GPU latest attempt to comply with strict export restrictions * WCCFTech: NVIDIA Preps GeForce RTX 5090 DD For China As Export-Compliant Model, Reportedly Features Blackwell GB202-240 GPU June 23: Reuters reported that DeepSeek is supporting China’s military and intelligence operations, based on an interview with a senior US State Department official. The official said DeepSeek was using “shell companies” in Southeast Asia to circumvent export restrictions. Reuters included comments from NVIDIA: “‘We do not support parties that have violated U.S. export controls or are on the U.S. entity lists,’ an NVIDIA spokesman said in a prepared statement, adding that ‘with the current export controls, we are effectively out of the China data center market, which is now served only by competitors such as Huawei.’” [...] “‘Our review indicates that DeepSeek used lawfully acquired H800 products, not H100,’ an NVIDIA spokesman said, responding to a Reuters query about DeepSeek's alleged usage of H100 chips.” DeepSeek did not respond to an inquiry from Reuters. * Sources: * Reuters: Exclusive: DeepSeek aids China's military and evaded export controls, US official says * Asia Times: DeepSeek gets NVIDIA’s high-end GPUs via Singapore: US official June 26: The Information reported that DeepSeek’s next AI model has been delayed due to a shortage of NVIDIA AI GPUs in China. * Sources: * The Information: DeepSeek’s Progress Stalled by U.S. Export Controls * Tom’s Hardware: AI disruptor DeepSeek's next-gen model delayed by NVIDIA GPU export restrictions to China — short supply of AI GPUs hinders development ##### July 2025 July 4: Bloomberg reported the US Department of Commerce is preparing a new export controls rule that would restrict the export of AI chips to Malaysia and Thailand. The rule’s goal would be to reduce AI chip smuggling to China. Based on its sources, Bloomberg said the export controls rule had not yet been finalized. * Sources: * Bloomberg: US Plans AI Chip Curbs on Malaysia, Thailand Over China Concerns July 10: Bloomberg reported that NVIDIA CEO Jensen Huang met with US President Donald Trump at the White House before traveling overseas to China. NVIDIA did not immediately respond to a request for comment from Reuters. * Sources: * Bloomberg: NVIDIA’s Jensen Huang Meets with Trump Ahead of CEO’s China Trip * Reuters: NVIDIA CEO Huang to meet Trump before China trip, source says * CNBC: Trump hosts Jensen Huang at White House as NVIDIA tops $4 trillion market cap July 11: In a public letter, a bipartisan group of US senators requested NVIDIA CEO Jensen Huang to avoid meeting with Chinese companies in an upcoming China trip that violate US laws or develop military applications that could undermine national security. Reuters included a response from NVIDIA about the senators’ letter: “An NVIDIA spokesperson said, ‘American wins’ when its technology sets ‘the global standard,’ and that China has one of the largest bodies of software developers in the world. AI software ‘should run best on the U.S. technology stack, encouraging nations worldwide to choose America,’ the spokesperson said.” * Sources: * United States Senate letter to Jensen Huang * The Hill: Bipartisan senators press NVIDIA CEO over China trip * Reuters: US senators warn NVIDIA CEO about upcoming China trip * International Business Times: NVIDIA CEO Huang to Face Chinese Officials Over AI Export Curbs Just as Company Touches $4 Trillion Milestone July 14: NVIDIA said it would soon resume sales of the H20 for customers in China. NVIDIA provided the following update in a blog post: “[Jensen] Huang also provided an update to customers, noting that NVIDIA is filing applications to sell the NVIDIA H20 GPU again. The U.S. government has assured NVIDIA that licenses will be granted, and NVIDIA hopes to start deliveries soon.” The White House did not respond to a request for comment from CNN. * Sources: * NVIDIA Blog: NVIDIA CEO Jensen Huang Promotes AI in Washington, DC and China * CNN: NVIDIA says it will restart sales of a key AI chip to China, in a reversal of US restrictions * Reuters: Chinese firms rush to buy NVIDIA AI chips as sales set to resume * Associated Press: NVIDIA to resume sales of highly desired AI computer chips to China The Malaysian government began requiring trade permits for all high-performance AI chips acquired from the United States. In a statement, the Malaysian government said: “The Ministry of Investment, Trade and Industry (MITI) would like to announce that, effective immediately, all exports, tranships and transits of high-performance AI chips of US origin are subject to a Strategic Trade Permit. These powers are provided for under Section 12 of the Strategic Trade Act 2010 (STA 2010), a Catch-All Control provision which requires individuals or companies to notify the relevant authority at least 30 days before exporting, transhipping, or bringing in transit any item not expressly listed in the Strategic Items List (SIL), if the individual or company knows or have reasonable grounds to suspect the item will be misused, or used for a restricted activity. This initiative serves to close regulatory gaps while Malaysia undertakes further review on the inclusion of high-performance AI chips of US origin into the SIL of the STA 2010. Malaysia stands firm against any attempt to circumvent export controls or engage in illicit trade activities by any individual or company, who will face strict legal action if found violating the STA 2010 or related laws.” * Sources: * Malaysia Government Release * Reuters: Malaysia says trade permit required for AI chips of U.S. origin * Bloomberg: Malaysia Controls AI Chip Exports As US Targets China Smuggling July 15: DigiTimes reported that NVIDIA is preparing a new AI GPU for the China market, the RTX 6000D. DigiTimes claimed the card would become available in the third quarter of 2025, according to its sources in the supply chain. * Sources: * DigiTimes: Exclusive: Jensen Huang's third visit to China in 2025; RTX 6000D aims for two million shipments * Tom’s Hardware: NVIDIA reportedly preparing RTX 6000D for Chinese market to comply with U.S. export controls — fabricated on TSMC N4, featuring GDDR7 memory capable of delivering 1,100 GB/s of bidirectional bandwidth July 24: The Financial Times reported that more than $1B worth of NVIDIA’s AI chips had been smuggled to China. In response, NVIDIA said that building datacenters with “smuggled products” was a “losing proposition.” * NVIDIA statement: * “Trying to cobble together datacenters from smuggled products is a losing proposition, both technically and economically. Datacenters require service and support, which we provide only to authorized NVIDIA products.” - NVIDIA to CNBC * Sources: * Financial Times: NVIDIA AI chips worth $1bn smuggled to China after Trump export controls * CNBC: NVIDIA addresses AI chip smuggling, says bootleg data centers are a ‘losing proposition’ July 28: The Financial Times reported that the US Commerce Department was not going to make “tough moves” to tighten export controls to China. According to the report, the US government would try to secure a better trade deal with China ahead of negotiations in Stockholm. The Washington Post reported that several congressional members had warned the US administration against loosening its export controls for AI GPUs. NVIDIA and the US Commerce Department did reply to requests for comment to The Washington Post. Several national security experts voiced their concern by sending a letter to the US Commerce Department. * Sources: * Financial Times: Donald Trump freezes export controls to secure trade deal with China * The Washington Post: Trump’s retreat on China chip ban triggers policy spat * Tom’s Hardware: Trump freeze on export restrictions to China reportedly in aid of trade talks — White House seeking face-to-face with Xi Jinping as dissenters warn H20 reversal is a dangerous mis-step July 29: Reuters reported that NVIDIA had ordered 300,000 more H20 chips from TSMC due to strong demand from its customers in China. Several weeks prior, NVIDIA said it would resume sales of the H20 chip to China. NVIDIA declined to comment to Reuters. * Sources: * Reuters: Exclusive: NVIDIA orders 300,000 H20 chips from TSMC due to robust China demand, sources say * Repost in MSN * Hot Hardware: TSMC Secures 300,000 H20 AI Chip Order As NVIDIA Boosts Supply To China July 31: The New York Times reported that Chinese government officials asked NVIDIA for information about security risks associated with its H20 chip. NVIDIA denied having “backdoors” in its chips. * NVIDIA statement: * “Cybersecurity is critically important to us. NVIDIA does not have ‘backdoors’ in our chips that would give anyone a remote way to access or control them.” - NVIDIA to CNBC * Sources: * The New York Times: China Summons NVIDIA Over ‘Backdoor Security’ Risks of A.I. Chips * CNBC: NVIDIA denies its China-bound H20 AI chips have ‘backdoors’ after Beijing’s security concerns * Reuters: NVIDIA says its chips have no 'backdoors' after China flags H20 security concerns * Associated Press: China summons NVIDIA over ‘backdoor safety risks’ in H20 chips ##### August 2025 Buy a GN 4-Pack of PC-themed 3D Coasters! These high-quality, durable, flexible coasters ship in a pack of 4, each with a fully custom design made by GN's team. You'll get a motherboard-themed coaster with debug display & reset buttons, a SATA SSD with to-scale connectors, RAM sticks, and a GN logo. These fund our web work! Buy here. August 4: A government official told Bloomberg the United States is exploring adding location trackers for AI chips. Bloomberg quoted the official as saying, “There is discussion about potentially the types of software or physical changes you could make to the chips themselves to do better location-tracking.” * Sources: * Bloomberg: US Explores Location Trackers for AI Chips, Official Says * The Register: Uncle Sam floats tracking tech to keep AI chips out of China August 5: In a blog post, NVIDIA said that its GPU products do not have backdoors or kill switches. * Sources: * NVIDIA Blog: No Backdoors. No Kill Switches. No Spyware. * CNBC: NVIDIA says its AI chips don’t have a ‘kill switch’ after Chinese accusation The US Department of Justice announced it had arrested two people in California for smuggling high-end GPUs to China that purportedly amount to “tens of millions of dollars’ worth of sensitive microchips used in artificial intelligence (AI) applications.” The BBC reported that court documents say the shipments included the NVIDIA H100 and RTX 4090. * NVIDIA statement: * “This case demonstrates that smuggling is a nonstarter. We primarily sell our products to well-known partners, including OEMs, who help us ensure that all sales comply with U.S. export control rules. Even relatively small exporters and shipments are subject to thorough review and scrutiny, and any diverted products would have no service, support, or updates.” - NVIDIA to TechCrunch * Sources: * Department of Justice release * Bloomberg: US Charges Chinese Nationals With NVIDIA Chips Export Breach * Reuters: Two Chinese nationals in California accused of illegally shipping NVIDIA AI chips to China * New York Post: Chinese nationals living in US charged with smuggling millions worth of NVIDIA’s powerful AI chips to Beijing * TechCrunch: Two arrested for smuggling AI chips to China; NVIDIA says no to kill switches * BBC: Chinese nationals charged with exporting NVIDIA AI chips to China August 10: The Financial Times reported that NVIDIA would give the US government 15% of its revenue from H20 chip sales from customers in China. The deal is reportedly part of an agreement that would allow NVIDIA to acquire export licenses from the Commerce Department in order to sell the H20 chip in China. AMD would be subject to the same rules for the MI308. * NVIDIA statement: * “We follow rules the U.S. government sets for our participation in worldwide markets. While we haven’t shipped H20 to China for months, we hope export control rules will let America compete in China and worldwide. America cannot repeat 5G and lose telecommunication leadership. America’s AI tech stack can be the world’s standard if we race.” - NVIDIA to the Associated Press * Sources: * Financial Times: NVIDIA and AMD to pay 15% of China chip sale revenues to US government * The Associated Press: NVIDIA and AMD to pay 15% of China chip sale revenue to US government in an unusual agreement * The New York Times: U.S. Government to Take Cut of NVIDIA and AMD A.I. Chip Sales to China * BBC: NVIDIA and AMD to pay 15% of China chip sales to US August 11: According to a report in Bloomberg, US President Trump said he was open to allowing NVIDIA to sell a modified Blackwell chip for the China market. The US President also said that he has negotiated with NVIDIA CEO Jensen Huang about the deal to allow H20 sales in China. * Sources: * Bloomberg: Trump Open to NVIDIA Selling Scaled-Back Blackwell Chip to China * The Register: Trump seeing green as he weighs deal to allow NVIDIA Blackwell GPU sales to China * NPR: Trump says NVIDIA will hand the U.S. 15% of its H20 chip sales to China * Reuters: Trump opens door to sales of version of NVIDIA’s next-gen AI chips in China * Repost in MSN August 12: Bloomberg reported that Chinese officials had “urged local companies” to avoid purchasing and using NVIDIA’s H20 chip, especially for national security and government work. According to Bloomberg, China questioned companies whether they had found security problems with NVIDIA’s chips. Bloomberg included commentary from NVIDIA: “AMD declined to comment, while NVIDIA said in a statement that ‘the H20 is not a military product or for government infrastructure.’ China has ample supplies of domestic chips, NVIDIA said, and ‘won’t and never has relied on American chips for government operations.’ * Sources: * Bloomberg: China Urges Firms Not to Use NVIDIA H20 Chips In New Guidance * Reuters: China cautions tech firms over NVIDIA H20 AI chip purchases, sources say CNBC reported that the Trump Administration was still working on the details for how to implement the 15% export tax on NVIDIA and AMD for selling certain chips to China. * NVIDIA statement: * “We follow rules the U.S. government sets for our participation in worldwide markets.” - NVIDIA to CNBC * Sources: * CNBC: White House says it’s working out legality of NVIDIA and AMD China chip deals * Tom’s Hardware: White House confirms it's still figuring out the legality of the revenue-sharing NVIDIA and AMD deal for China GPU sales — 'The legality of it, the mechanics of it, is still being ironed out' August 13: Reuters reported that US officials have covertly placed “location-tracking devices” in targeted shipments with advanced chips in an effort to catch chip smuggling to China. Unnamed sources told Reuters that the tracking devices had been placed in shipments of OEM servers, including from Dell and Supermicro. NVIDIA declined to comment to Reuters. * Sources: * Reuters: Exclusive: US embeds trackers in AI chip shipments to catch diversions to China, sources say * Tom’s Hardware: U.S. authorities allegedly placed secret tracking devices in AI chip shipments to China — report claims targeted shipments from Dell and Super Micro containing NVIDIA and AMD chips had trackers in packaging and servers themselves TweakTown: US authorities secretly place location tracking devices in targeted AI chip shipments to China * * *