While [1]down to AMD Austin yesterday for the Instinct MI355X and [2]ROCm 7.0 launch, I had the chance to chat again with Anush Elangovan. As the VP of AI Software at AMD, talking with Anush is always insightful and technical in nature. One of the questions I posed him was around the length of hardware support with ROCm.

Yes, the official ROCm hardware support matrix for Linux is very sparse on the consumer side. But, yes, there are many more GPUs in effect supported by ROCm than what makes the official support matrix list. The [3]ROCm system requirements with the Radeon GPU supported list is well short of all the expanding list of AMD GPUs that can successfully run with ROCm but it comes down to what is officially supported/validated. Especially if diving into [4]TheRock for enjoying the latest ROCm bits and getting ROCm working on "unsupported" GPUs.

So while the support matrix isn't comprehensive for everywhere that ROCm can actually run, my question was focused on when AMD will phase out "old" hardware support. As it's one thing to not officially support/validate older GPUs but to not actively remove the support or anything to regress that older support intentionally.

The perfect example came up yesterday with [5]the ROCm 7.0 release . The [6]7.0 release notes where it mentions the ROCm Compute Profiler has " Removed support for AMD Instinct MI50 and MI60. " The AMD Instinct MI50 was already marked as "unsupported" by the official support matrix, but is now AMD actively removing support? The MI50 and MI60 were launched back in 2018 as a reminder.

Anush wasn't aware of the ROCm profiler removing the MI50/MI60 support. But he said he would be looking into it for something that could potentially be reverted.

Indeed when digging into it after the fact, [7]this commit in July dropped the MI50/MI60 specific data from the profiler. This would be straight-forward to revert and carries little technical debt with just being a number of YAML files.

He would ideally like to see AMD hardware "supported" by ROCm for around ten years. But, of course, there are caveats and at least the 6~7 year period is reasonable if not being able to hit ten years. But at AMD they wouldn't actively be working to remove old hardware support without cause.

With the AMDGPU/AMDKFD kernel drivers being open-source and upstream in the Linux kernel they continue to carry hardware support long past AMD's official hardware support periods. With ROCm being open-source too, at least there is the transparency into these changes and the possibility of reverting changes made by AMD / community forks, etc.

Long story short, it sounds like AMD intends to keep GPUs "supported" by ROCm at least as long as NVIDIA hardware by CUDA if not longer. I was comfortable with his answers and AMD's plans around long-term hardware support. Plus being open-source the community can continue doing their thing long past when AMD actively ends supporting a given product. As for the MI50/MI60 profiler change specifically, Anush is looking into it.

And a fun piece of trivia that came up in the discussions and for the Arch vs. Gentoo passionate (vocal) users in the forums... Anush is a die-hard Gentoo Linux user since 2001. Thanks for AMD having me out in Austin to discuss the ROCm 7.0 release.



[1] https://www.phoronix.com/news/AMD-Instinct-MI355X-ROCm-7.0

[2] https://www.phoronix.com/search/ROCm+7.0

[3] https://rocm.docs.amd.com/projects/install-on-linux/en/latest/reference/system-requirements.html

[4] https://github.com/ROCm/TheRock

[5] https://www.phoronix.com/news/AMD-ROCm-7.0-Released

[6] https://rocm.docs.amd.com/en/latest/about/release-notes.html

[7] https://github.com/ROCm/rocm-systems/commit/537a269e953eb851e439021fed0efcaad39357a3