AMD ROCm 5.4.2 Released As Another Small Update To The Compute Stack

[phoronix]

Phoronix: AMD ROCm 5.4.2 Released As Another Small Update To The Compute Stack

ROCm 5.4 released in November with a point release then coming out in December and now there is another minor update for January to this open-source AMD Linux GPU compute stack...

https://www.phoronix.com/news/AMD-ROCm-5.4.2

[Volker Schmidt]

Sometimes I wonder, whether hardware like Radeon Instinct MI50/MI60 is really that overpriced, making that much money, that they are worth ignoring consumers, who simply wish to be able to some computing, for fun, education or testing on there consumer graphic cards. Leaving them with no other choice but to buy NVIDIA?

[littlewu2508]

https://github.com/ROCmSoftwarePlatform/MIOpen/commit/230a8f9322548f9ae3a879711a682f6bfe00e2ab

MIOpen finally kicks off its journey towards gfx11!

[dreamcat4]

The officially supported hardware by ROCm remains the Radeon Instinct MI50/MI60, Radeon VII, Radeon Pro VII, Radeon Pro W6800 series, Radeon Pro V620, and Instinct MI100/MI200 hardware. There's nothing to report yet on official RDNA3 support for ROCm.

amd are being really slow in updating their documentation. which just isn't up to scratch to give any good representation what actually works in reality or not. and with which specific compute libraries or features. this has greatly discouraged many developers from even bothering to look at providing amd compute support. as its largely assumed not to be working... which is mostly not the case. but heh amd didnt get the memo (or their own memos). to know how important to tell us that

however there has been a long time a tracking issue on github about it here

https://github.com/RadeonOpenCompute/ROCm/issues/1714

and coming from that issue a community member has been spending a lot of time trying to submit a table for amd to publish in their official docs (which atto still hasnt happened yet, despite being mostly complete). it seems that being perfectionists they couldn't think to just release it by now with draft written on it. and some simple * disclaimer warning text to know that some details might be incorrect or subject to change. but that's too easy. gotta make up some other superficial barrier to not tell the world until its far too late. anyhow heres the table in un-release pr form. might require some downloading of files in git. and rendering as html tables or whatever the formatting its in.

https://github.com/RadeonOpenCompute/ROCm/pull/1738

[evasb]

I'm not of the field. But is AMD stack that bad compared with NVIDIA?

[bridgman]

Sometimes I wonder, whether hardware like Radeon Instinct MI50/MI60 is really that overpriced, making that much money, that they are worth ignoring consumers, who simply wish to be able to some computing, for fun, education or testing on there consumer graphic cards. Leaving them with no other choice but to buy NVIDIA?

If something runs on MI50/60 it will run on the consumer equivalent (Radeon VII) as well. We don't disable consumer SKUs.

What the MI50/60 adds is more memory, full ECC and XGMI links for connecting GPUs together.

AFAIK the complaint is not about support for consumer equivalents to Instinct cards but rather about delays in support for consumer generations that had no data center equivalent, particularly RDNA1 and RDNA2.

[_r00t-]

RustiCL FTW!

[qarium]

If something runs on MI50/60 it will run on the consumer equivalent (Radeon VII) as well. We don't disable consumer SKUs.
What the MI50/60 adds is more memory, full ECC and XGMI links for connecting GPUs together.
AFAIK the complaint is not about support for consumer equivalents to Instinct cards but rather about delays in support for consumer generations that had no data center equivalent, particularly RDNA1 and RDNA2.

my vega64 still does not work in fedora 37 ... maybe vega64 is EOL end of life and i should upgrade... but last time i checked blender and ROCm added vega64 support...

maybe amd should do make packages for distros like fedora.

[wsippel]

If something runs on MI50/60 it will run on the consumer equivalent (Radeon VII) as well. We don't disable consumer SKUs.

What the MI50/60 adds is more memory, full ECC and XGMI links for connecting GPUs together.

AFAIK the complaint is not about support for consumer equivalents to Instinct cards but rather about delays in support for consumer generations that had no data center equivalent, particularly RDNA1 and RDNA2.

Or RDNA3. Which is really sad, considering it's AMD's first consumer GPU with AI accelerators - AI accelerators you won't be able to use until at least ROCm 5.5, which doesn't even have a tentative release window. Pytorch and Tensorflow don't work at all on RDNA3 with ROCm 5.4, not even using the RDNA2 code path. But we did get day one support for HIP this time, which is a step in the right direction and much appreciated.

[finalzone]

my vega64 still does not work in fedora 37 ... maybe vega64 is EOL end of life and i should upgrade... but last time i checked blender and ROCm added vega64 support...
maybe amd should do make packages for distros like fedora.

Strange issue with your Vega64. I am running on AMD Ryzen 7 5825U which has a mobile Vega8 (named as AMD Radeon Graphic with rocm-opencl. Only rocm-hip-runtime has issue yet runs fine on Radeon 6950XT.