AMD Details The MI300X & MI300A, Announces ROCm 6.0 Software

[Svyatko]

I totally understand they couldn't clone or copy CUDA. But, they have had the OpenCL 'problem' for years, apparently. I've read some decent 'rants' on the OpenCL/AMD thing - but, I just can't remember all of it. All I know is it's going back many years - at least 5 or more.

https://www.reddit.com/r/Amd/comments/bmyux8/opencl_no_longer_supported_by_amd/

https://community.amd.com/t5/drivers-software/opencl-where-is-it-has-amd-removed-it-quietly/td-p/410595

https://www.reddit.com/r/Amd/comments/mtz22q/who_is_to_blame_for_the_bad_opencl_performance/

I understand it's been problematic or 'broken' for years and the investment/priority has been extremely lacking - and the lack of focus or addressing the problem has been demonstrated in fields like Blender and other areas that require Compute (and the OpenCL backend). They switched to HIP and that looks like a similar situation - perhaps, a little bit of progress but REALLY slow and mediocre performance when using it - and anything related - i.e. ray tracing library.

What AMD really lacks are brains in managers heads.
I've managed to use AMD PAL OpenCL driver on Linux with AMD APU, which AMD just has dropped.
I did it by myself. This was not magic, this was hard work.
Why I can do this, but AMD cannot?

Here are my results: https://en.opensuse.org/SDB:AMD_GPGPU

My point, before - is that HIP is still a WIP and they still haven't fixed the RT aka ray tracing acceleration in Blender - which would be the ONLY reason to even consider an AMD card for that. Nvidia is evil but most ppl still use it for whatever tasks - in these fields - GPU COMPUTE, Blender, ML/AI etc. etc. - the only reason ppl are buying AMD cards right now is the price is a bit less and in the Linux world - for gaming - since, there's (supposedly) less issues to deal with - but, it's my impression that it's mostly ppl who don't do much with their PC or haven't done much in Linux other than gaming - I guess the community here has a lot of those?

Because, when you try to do more than just gaming in Linux - you encounter stress and problems - ROCm, getting that configured w/ HIP is often a hassle - and it's my understanding that OpenGL can be implemented with open source components but NOT OpenCL - which is where AMD users get into trouble - trying to 'add' proprietary components into the FOSS world of their AMD software/files.... but, correct me if I'm mistaken there? I don't think so....

Open-source OpenCL drivers: ROCm, Rusticl.
AMD is not using dedicated CUs for RT - a contraria to Nvidia & Intel. That causes HIP RT with Blender having small advantages comparing to ordinary HIP - from minus 10% to plus 20%, about +5% in average. Do you want +5% to speed with adding some glitches?

For now, I'm keeping an eye on how the 7900 series shapes up in being a potential option in these specific areas - or should I just hunt for a used 3090 or something?

ROCm 5.7 supports RDNA3 only with Ubuntu 22.04.
ROCm 6.0 possible will support other distributions.

Intel Arc is good in Blender. Possibly Intel 2nd gen GPUs - Battlemage - will be better.

https://opendata.blender.org/benchmarks/query/?device_name=arc&blender_version=3.6.0&group_by=device_name&group_by=operating_system&group_by=compute_type

[Panix]

What AMD really lacks are brains in managers heads.
I've managed to use AMD PAL OpenCL driver on Linux with AMD APU, which AMD just have dropped.
I did it by myself. This was not magic, this was hard work.
Why I can do this, but AMD cannot?

Here are my results: https://en.opensuse.org/SDB:AMD_GPGPU




Open-source OpenCL drivers: ROCm, Rusticl.
AMD is not using dedicated CUs for RT - a contraria to Nvidia & Intel. That causes HIP RT with Blender having small advantages comparing to ordinary HIP - from minus 10% to plus 20%, about +5% in average. Do you want +5% to speed with adding some glitches?




ROCm 5.7 supports RDNA3 only with Ubuntu 22.04.
ROCm 6.0 possible will support other distributions.

Intel Arc is good in Blender. Possibly Intel 2nd gen GPUs - Battlemage - will be better.
https://opendata.blender.org/benchmarks/query/?device_name=arc&blender_version=3.6.0&group_by=de vice_name&group_by=operating_system&group_by=compu te_type

It was my impression, that HIP-RT *could* work surprisingly well on 7900 series cards - up to 25% well on some Cycles - depending on variables? But, if you are implying, the improvement on those gpus are only modest, at best - then that is even more damning.

As for ROCm - yeah, I looked at those official doc. pages - the Windows support appears to be superior - and the support in Linux - is limited to only a few cards. It looks pretty awful, imho.

https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.0.0/reference/system-requirements.html

Look at this - a year ppl were complaining - and their progress is.....hardly anything....

https://github.com/ROCm/ROCm/issues/1874

https://www.forbes.com/sites/karlfreund/2023/12/13/breaking-amd-is-not-the-fastest-gpu-heres-the-real-data/?sh=64db064c3a6f

These are corporate liars - both will claim they are 'better' than the other - but, there's no reason to be a fanboi of either company - wait for 3rd party testers to compare the technologies and hardware - and even then, that's difficult from what I'm reading - with their AI implementations - I wouldn't trust either - but, AMD's vows of 'great support and performance' has a history of being inaccurate - and this is probably another one.

[Bane]

A couple of notes:

I force clang for building all the ROCm components at the moment because it makes it easier to deal with stripping unsupported flags which trip up the HIP compiler.

I also have ebuilds for flang(-new) with a patch to rename it to "flang" and some of the component ebuilds are hard set to use it. I wasn't going to leave it like that when I planned to push, but it's done now... I have this in this overlay because it makes it possible to write fortran code targeting amdgcn.

The mlir ebuild (needed for flang) is WIP. It needs fixing up for the various useflags. It builds with how I have mine set, no promises it will work for you. Contributions welcome! ;-)

Thanks for this