Testing ROCm on ARM64 and PPC64

AMD only supports ROCm on x86-64, but the Debian AI Team has patched the ROCm libraries so that they compile on ARM64 and PPC64. Debian chose these additional build targets on the basis that the Kernel Fusion Driver (KFD) used for ROCm was enabled on ARM64 and PPC64 (in Linux commits d1c234e and c38402f). However, nobody in the Debian AI community seems to have access to ARM64 or PPC64 machines with PCIe slots, so nobody has ever tested if Debian’s ROCm packages work on those platforms.

Does anybody in the Level1Techs community have an appropriate ARM64 system and an old AMD GPU like the Vega 64 or MI25 lying around? Debian has packaged the unit tests for most of the ROCm math libraries and it would be interesting to run those unit tests on ARM or PPC systems just to see what happens.

The first test would just be to check that the driver initialized successfully and that rocminfo works.

2 Likes

I’ve got a Snapdragon X Elite, I’ll definitely test this out but we’re talking months before USB4 starts working properly on this hw I think. it’d be best if someone with a rpi4/5 chimed in but I think this is probably one of the nichest use cases ever lol.

I appreciate the offer, but I don’t know whether ROCm works over USB even on x86-64, so that may be introducing some additional points of failure. If it doesn’t work (which is likely), we still won’t know exactly why.

For ARM systems, I’m hoping someone has a system with a working NVIDIA GPU that they could swap out temporarily. An example of such a system would be the Gigabyte G242-P33.

Raptor Computing has offered to run some tests on PPC64, so I think that architecture will soon be tested.

Are you still looking for someone to test ARM64? I have a homelab project I just started working on. I’m going to try building a local LLM server on an Ampere Altra Dev Kit with a pair of AMD Radeon Pro v620’s to run Ollama and OpenWebUI. I decided to give it a shot after I’d heard about the ROCm drivers coming to ARM64. I’d be willing to help with a few basic tests once I get all my parts in and provided I can get everything working.

I have a setup of ARM Neoverse cpu connected via PCI to MI210.
I want to run llm on it, preferably with ROCm.
I already handled the driver part and amdgpu pass probe ok.
Let’s connect and I will be happy to test your ROCm on ARM.

I can’t help here, but just wanted to voice support for this project. Thanks for doing this!