Got ComfyUI running. The various elements that have to talk to each other for this work are modestly involved:
GPU, Strix Halo (gfx1151)
OS, Ubuntu 25.04
Python, 3.13.7
ROCM, 7.1
Git
Pytorch -- the hardest thing to get running, it required a nightly build that it took me serveral hours to find, […]
GPU, Strix Halo (gfx1151)
OS, Ubuntu 25.04
Python, 3.13.7
ROCM, 7.1
Git
Pytorch -- the hardest thing to get running, it required a nightly build that it took me serveral hours to find, […]
Original post on mstdn.social
mstdn.social
November 13, 2025 at 2:47 AM
Got ComfyUI running. The various elements that have to talk to each other for this work are modestly involved:
GPU, Strix Halo (gfx1151)
OS, Ubuntu 25.04
Python, 3.13.7
ROCM, 7.1
Git
Pytorch -- the hardest thing to get running, it required a nightly build that it took me serveral hours to find, […]
GPU, Strix Halo (gfx1151)
OS, Ubuntu 25.04
Python, 3.13.7
ROCM, 7.1
Git
Pytorch -- the hardest thing to get running, it required a nightly build that it took me serveral hours to find, […]
I agree the CUDA stack has a lot of lock-in and decades of investment, but I also have seen a lot of large companies and communities investing time to make rocm stack better because the hardware is so much cheaper and available
November 12, 2025 at 12:56 PM
I agree the CUDA stack has a lot of lock-in and decades of investment, but I also have seen a lot of large companies and communities investing time to make rocm stack better because the hardware is so much cheaper and available
📊 Multi-sourcing by hyperscalers supports AMD’s case; any sign of faster ROCm adoption and developer traction would be a catalyst.
- Supply chain beneficiaries: Tightness in HBM and advanced packaging continues. Watch TSMC CoWoS capacity and memory suppliers (HBM) for knock-on demand. (7/10)
- Supply chain beneficiaries: Tightness in HBM and advanced packaging continues. Watch TSMC CoWoS capacity and memory suppliers (HBM) for knock-on demand. (7/10)
November 12, 2025 at 12:45 PM
📊 Multi-sourcing by hyperscalers supports AMD’s case; any sign of faster ROCm adoption and developer traction would be a catalyst.
- Supply chain beneficiaries: Tightness in HBM and advanced packaging continues. Watch TSMC CoWoS capacity and memory suppliers (HBM) for knock-on demand. (7/10)
- Supply chain beneficiaries: Tightness in HBM and advanced packaging continues. Watch TSMC CoWoS capacity and memory suppliers (HBM) for knock-on demand. (7/10)
I'd be concerned about rocm vs cuda? Cuda being so much more established.
Last I heard someone try to do LLMs with rocm they were very sad and frustrated but this is probably 1.5 years back. Hopefully things are better now. If I explore this stuff I'd much prefer it to be AMD.
Last I heard someone try to do LLMs with rocm they were very sad and frustrated but this is probably 1.5 years back. Hopefully things are better now. If I explore this stuff I'd much prefer it to be AMD.
November 12, 2025 at 12:11 PM
I'd be concerned about rocm vs cuda? Cuda being so much more established.
Last I heard someone try to do LLMs with rocm they were very sad and frustrated but this is probably 1.5 years back. Hopefully things are better now. If I explore this stuff I'd much prefer it to be AMD.
Last I heard someone try to do LLMs with rocm they were very sad and frustrated but this is probably 1.5 years back. Hopefully things are better now. If I explore this stuff I'd much prefer it to be AMD.
November 12, 2025 at 8:40 AM
puzzling cmake error
put it into google
google tells you you're an idiot and HIP's great actually
put it into google
google tells you you're an idiot and HIP's great actually
![Google search: "No known features for HIP compiler"
• Al Overview The statement "no known features for HIP compiler" is incorrect. The HIP compiler is a key component of the AMD ROCm platform, enabling developers to write code for both AMD and NVIDIA GPUs. It supports features like offline and runtime compilation, porting from CUDA, and C++ language extensions. A potential reason for this user's confusion could be issues with cross-platform compatibility or configuration, which are sometimes challenging to set up correctly. e HIP compiler features • Source portability: HIP allows you to write code that can be compiled and run on different hardware, including AMD GPUs and NVIDIA GPUs, though not as a single)
Frequently asked questions — HIP Documentation Can a HIP binary run on both AMD and Nvidia platforms? # HIP is a source-portable language that can be compiled to r... AMD ROCm documentation : [Issue]: no cross-platform compilation #3582 - ROCm/hip - GitHub 22 Aug 2024 — Problem Description... This feature seems to be badly documented and...
Show more v](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:gtpogdkd4kyv6yuy7qcb747o/bafkreibo3ra3ulepq7outvnmxcshjtkkfov72qlitqovg5sxeeagk4uexu@jpeg)
November 11, 2025 at 10:44 PM
puzzling cmake error
put it into google
google tells you you're an idiot and HIP's great actually
put it into google
google tells you you're an idiot and HIP's great actually
AMD was definitely behind in software with ROCM and has taken catching up very seriously.
November 11, 2025 at 6:32 PM
AMD was definitely behind in software with ROCM and has taken catching up very seriously.
winbuzzer.com/2025/11/11/m...
Microsoft Apparently Wants to “Break” Nvidia’s Moat, Making CUDA Available to AMD AI Chips
#AI #Microsoft #NVIDIA #AMD #CUDA #CloudComputing #ROCm #AIInfrastructure #GPU #BigTech #Semiconductors
Microsoft Apparently Wants to “Break” Nvidia’s Moat, Making CUDA Available to AMD AI Chips
#AI #Microsoft #NVIDIA #AMD #CUDA #CloudComputing #ROCm #AIInfrastructure #GPU #BigTech #Semiconductors
Microsoft Apparently Wants to “Break” Nvidia’s Moat, Making CUDA Available to AMD AI Chips - WinBuzzer
Microsoft appears to be developing a toolkit to convert NVIDIA CUDA models to run on AMD's ROCm platform, a move aimed at cutting AI inference costs and challenging NVIDIA's software dominance.
winbuzzer.com
November 11, 2025 at 2:05 PM
winbuzzer.com/2025/11/11/m...
Microsoft Apparently Wants to “Break” Nvidia’s Moat, Making CUDA Available to AMD AI Chips
#AI #Microsoft #NVIDIA #AMD #CUDA #CloudComputing #ROCm #AIInfrastructure #GPU #BigTech #Semiconductors
Microsoft Apparently Wants to “Break” Nvidia’s Moat, Making CUDA Available to AMD AI Chips
#AI #Microsoft #NVIDIA #AMD #CUDA #CloudComputing #ROCm #AIInfrastructure #GPU #BigTech #Semiconductors
AI: Interesting development.
NVidia has had a decade or more of market monopoly with CUDA. If Micro$oft pulls this off (port your CUDA code to RocM), then AMD is back in the game.
Which is GOOD for competition.
NVidia has had a decade or more of market monopoly with CUDA. If Micro$oft pulls this off (port your CUDA code to RocM), then AMD is back in the game.
Which is GOOD for competition.
Microsoft Code Allows CUDA to Work On AMD GPU's - MS Destroying NVIDIA
YouTube video by Eli the Computer Guy
www.youtube.com
November 11, 2025 at 1:31 PM
AI: Interesting development.
NVidia has had a decade or more of market monopoly with CUDA. If Micro$oft pulls this off (port your CUDA code to RocM), then AMD is back in the game.
Which is GOOD for competition.
NVidia has had a decade or more of market monopoly with CUDA. If Micro$oft pulls this off (port your CUDA code to RocM), then AMD is back in the game.
Which is GOOD for competition.
November 11, 2025 at 10:38 AM
November 11, 2025 at 1:10 AM
that AMD Ryzen Max+ 395 remains one of my favorite computer purchases in a very long time and that's saying a lot considering the state of doing anything with rocm. 96GB is assigned to the GPU leaving 32GB for a really great build machine for coding, so I'm distracted from doing models on it lol
November 10, 2025 at 4:48 PM
that AMD Ryzen Max+ 395 remains one of my favorite computer purchases in a very long time and that's saying a lot considering the state of doing anything with rocm. 96GB is assigned to the GPU leaving 32GB for a really great build machine for coding, so I'm distracted from doing models on it lol
November 10, 2025 at 2:04 PM
November 10, 2025 at 1:31 PM
AMD's GPUS for AI: Radeon, Instinct, and ROCm
YouTube video by AIProgrammingHardware
youtu.be
November 10, 2025 at 11:17 AM
November 10, 2025 at 4:56 AM
AMD's ROCm 7 And Helios Narrow The NVIDIA Gap
AMD's ROCm 7 And Helios Narrow The NVIDIA Gap
AMD's ROCm 7 And Helios Narrow The NVIDIA Gap
AMD's ROCm 7 And Helios Narrow The NVIDIA Gap
AMD's ROCm 7 And Helios Narrow The NVIDIA Gap
animeaura.store
November 9, 2025 at 2:50 PM
AMD's ROCm 7 And Helios Narrow The NVIDIA Gap
AMD's ROCm 7 And Helios Narrow The NVIDIA Gap
AMD's ROCm 7 And Helios Narrow The NVIDIA Gap
리눅스 랩탑(amdgpu, rocm)에서 ollama 써서 아주 잠깐 로컬 LLM 시도해 본 소감: 생각보다 잘 되는데, 한편으로 왜 OpenAI 의 적자가 저렇게 터무니없는 규모라는 것인지도 이해했다. 어디가 병목인지는 알겠는데 그래서 그에 대한 대응책은 GPU 를 증설하기 시작해서 그냥 많이 증설했습니다. 를 하고 있다는 것이군 진짜로… 이게 인류의 갈 길이라니… 사이버펑크보다 현실이 더 사이버펑크다
November 9, 2025 at 5:16 AM
리눅스 랩탑(amdgpu, rocm)에서 ollama 써서 아주 잠깐 로컬 LLM 시도해 본 소감: 생각보다 잘 되는데, 한편으로 왜 OpenAI 의 적자가 저렇게 터무니없는 규모라는 것인지도 이해했다. 어디가 병목인지는 알겠는데 그래서 그에 대한 대응책은 GPU 를 증설하기 시작해서 그냥 많이 증설했습니다. 를 하고 있다는 것이군 진짜로… 이게 인류의 갈 길이라니… 사이버펑크보다 현실이 더 사이버펑크다
Microsoft has developed internal toolkits that convert NVIDIA CUDA models into ROCm-compatible code to run on AMD AI GPUs (examples: 300X) and is working with AMD on 400X and 450X to reduce inference costs and cut NVIDIA CUDA lock-in.
Microsoft Is Building a 'CUDA Killer' to Break NVIDIA's AI Monopoly
Microsoft is exploring ways to leverage the 'stack' of its AMD GPUs for workloads, as it develops toolkits that convert CUDA into ROCm code.
wccftech.com
November 9, 2025 at 1:57 AM
Microsoft has developed internal toolkits that convert NVIDIA CUDA models into ROCm-compatible code to run on AMD AI GPUs (examples: 300X) and is working with AMD on 400X and 450X to reduce inference costs and cut NVIDIA CUDA lock-in.
November 7, 2025 at 9:52 PM
November 7, 2025 at 2:40 PM
November 7, 2025 at 9:49 AM
onnxruntime-rocm 1.22.2-12 x86_64 Cross-platform, high performance scoring engine for ML models (with ROCm)
Interest | Match | Feed
Interest | Match | Feed
Origin
archlinux.org
November 7, 2025 at 7:16 AM
onnxruntime-rocm 1.22.2-12 x86_64 Cross-platform, high performance scoring engine for ML models (with ROCm)
Interest | Match | Feed
Interest | Match | Feed
Origin
archlinux.org
November 7, 2025 at 7:16 AM