A. Schilling
aschilling.bsky.social
A. Schilling
@aschilling.bsky.social
630 followers 58 following 400 posts
Editor at hardwareluxx.de
Posts Media Videos Starter Packs
There is a company making billions per quarter, but doing an increasingly bad job in terms of PR. They are giving a lot of infos, also the (in my opinion) important ones, but are holding back on basic stuff like photos of these announcements or are not providing them willingly.
AMD started the scheme with their announcement a day before with the DoE, ORNL and Discovery announcement. Its almost like the AI hardware (and Oracle, HPE, ...) companies need those incentives to "sell" these systems to the research centers.
This is #Discovery, the next-gen supercomputer run by ORNL and build by HPE. Discovery will use AMD Instinct MI430X accelerators and Epyc "Venice" CPUs.

www.hardwareluxx.de/index.php/ne...
Single die and dual die implementation of the Maverick-2 by @nextsilicon.com. Single die package is used with the PCIe card while die dual die package sits on the OAM module.
Die shot from @nextsilicon.com. But this can't be the Maverick-2, because the aspect ratios look a bit strange, or rather, the HBM PHYs are positioned on the long side here. Or am I wrong? It's been a long week, so apologies in advance!
I‘m kind of shocked that people within the business thought that TSMC manufacturing wafer with Blackwell GPUs in Arizona are giving NVIDIA the ability to do all the silicon related manufacturing independent from Taiwan 🇹🇼
The two Broadcom #Tomahawk 6 102.4T switch chip in two package variants:

Tomahawk 6 112G SerDes (bigger package)
Tomahawk 6 224G SerDes (smaller package)

Source: @servethehome.com.web.brid.gy
"Measurements from 600 transistors showed low SS values and high channel current ON/OFF ratios, as well as high and similar field-effect mobilities across the stacks.

(...) but extending the thermal stability beyond 50 °C remains an important goal for future research (...)"
Researchers at KAUST have stacked six transistors (OxTs and OrTs in an alternating fashion) on top of each other. These stacks consist of 41 layers, fabricated using a 40-step lithography process, each containing 100 devices.
OpenAI and Tachyum announce strategic partnership to deploy 10 Quadrillion Watts of Compute Power without deploying any actual Chip made by Tachyum.
TSMC Revenue by Node as of Q3 2025.

We should see N2 emerging by Q1 2026 in this graph.
Apple M5

- 4x Performance Core
- 6x Efficiency Core
- +15% nT Performance
- +45% higher GPU Performance
- 153 GB/s Memory Interface
- 3 nm Process
Intel just announced Crescent Island. Crescent Island is a new Intel Data Center GPU offering that will arrive in the second half of 2026. Knows features:

- air-cooled
- Xe3P microarchitecture
- 160 GB of LPDDR5X

Looks like a competitor to Rubin CPX and token optimized XPUs.
At 2025 OCP Global Summit AMD shows their Helios rack-scale solution. These are liquid-cooled, 72-GPU racks powering AMD Instinct MI450 Series GPUs, AMD EPYC CPUs (Venice) and do feature integrated UALoE scale-up connectivity.
Interesting project by der8auer: Spray cooling with Novec 7000. Acceptable idle temperatures, nothing more. Just a POC.

But, spray cooling was used with the Cray X1 supercomputer. Small holes in the heatspreader sprayed directly onto the small chips, and the gas was returned.
It's official! OpenAI and Broadcom announce strategic collaboration to deploy 10 GW of OpenAI-designed AI accelerators.

OpenAI and Broadcom have signed a term sheet to deploy racks incorporating the AI accelerators and Broadcom networking solutions.
One of the slides that shows the potential of #PantherLake the best. 4P+8E+4LPE cores on par with 4P+4LPE (Lunar Lake) and better than 6P+8E+2LPE (Arrow Lake-H), where the allocation of those LPE cores had it’s hurdles.
Robinson Lake reference design for embedded/robotics #PantherLake solutions. The CPU is rated for an extended temperature range of -40 °C to 105 °C. The memory chips are labeled with 4ZC42 DBFVB from Micron. #IntelTechTour
This is a #Intel18A wafer with #PantherLake compute tiles.

I did a quick annotation (see 2nd picture attached).
I had the opportunity to visit Intel's #Fab52 at the Ocotillo campus in Chandler, Arizona. This is where Intel manufactures its chips in #Intel18A. In this report, I attempt to convey the enormous dimensions and challenges involved. #IntelTechTour

www.hardwareluxx.de/index.php/ne...
A switch from #Crestmont to #Darkmont and thus a significant leap in CPU performance, but still platform-compatible and technically impressive with #Intel18A + #FoverosDirect3D. Intel provides further details on #ClearwaterForest, also known as #Xeon6+.

www.hardwareluxx.de/index.php/ne...