AMD ROCm 6.1.2 Released With Fixes & Optimizations

[phoronix]

Phoronix: AMD ROCm 6.1.2 Released With Fixes & Optimizations

ROCm 6.1.2 is out today as the newest update to AMD's open-source GPU compute stack for Linux systems and with growing support for Windows Subsystem for Linux...

https://www.phoronix.com/news/AMD-ROCm-6.1.2-Released

[Dukenukemx]

Great, so how do I use this on my Vega 56 without screwing over my Kisak PPA drivers?

[NeoMorpheus]

Grabs the popcorn and sit on the porch to welcome the AMD haters

[djart]

Still no support for all RDNA 1/2/3 consumer GPUs, still no support for Ubuntu 24.04 and possibly other distros. Tsk, AMD you are moving too slow, if at all ....

[aerospace]

Generic Radeon hardware support please, generic Radeon hardware support please, generic Radeon hardware support please...

[danger89]

I'm happy with the increasing support and bug fixes in rocm. I hope amd will also focus on their gpu drivers as well.

Making AMD much more compelling then Nvidia in terms of costs vs performance.

[timofonic]

Still no support for all RDNA 1/2/3 consumer GPUs, still no support for Ubuntu 24.04 and possibly other distros. Tsk, AMD you are moving too slow, if at all ....

I agree with you and aerospace too!

AMD needs to wake up , invert faster and more in software engineering. CUDA is far from perfect, but they made CUDA available on all their GPUs and actively promoted it.

I think the future is Vulkan replacing OpenCL and others plus massive standardization and improvements against the green ones. Also, they should make something with OpenAL and probably merging into Vulkan. GPU hardware is quite capable of processing sound.

[numasan]

I have mixed emotions about ROCm and AMD's efforts, mostly disappointed unfortunately. Been testing out an RX 7900 XTX for compute, and coming from Nvidia the experience is very unpolished and doesn't live up to my expectations.

Minor points:
Why do they hardcode the ROCm point-release in the repo file? There should be no need to manually change from 6.1.1 to 6.1.2 - I'd expect a general system update to include this ROCm update.

Why does rocm-smi not show which ROCm version is used? The amd-smi tool does (is this the eventual replacement for rocm-smi?) but neither tool shows the GPU marketing name, best case is rocm-smi showing an unhelpful ID (0x744c,1341).

Major issue:
The performance is bad, really shockingly bad. In the PyTorch example "word_language_model" running this command "python main.py --cuda --emsize 1500 --nhid 1500 --dropout 0.65 --epochs 40 --tied" this are the average epoch times I get:

RX 7900 XTX ROCm = 218 seconds
RTX 2080 Ti CUDA = 183 seconds
RTX A5000 CUDA = 104 seconds
RTX 4090 CUDA = 53 seconds

How can a 5 year old 2080 Ti beats the current AMD consumer flagship?! Why should I ever consider an expensive AMD compute card if the performance is just as bad relative to the competition? It may be half price, but so is the performance apparently! Are the AMD consumer card severely limited in compute on purpose? Or, as I suspect, is ROCm just holding the hardware back?

[lu_tze]

Why do they hardcode the ROCm point-release in the repo file? There should be no need to manually change from 6.1.1 to 6.1.2 - I'd expect a general system update to include this ROCm update.

There's also "latest" as a version, that could do what you want. But are you sure you want it? AMD can and will break stuff between minor versions.

The performance is bad, really shockingly bad.

It runs and doesn't crash? Lucky you.

How can a 5 year old 2080 Ti beats the current AMD consumer flagship?! Why should I ever consider an expensive AMD compute card if the performance is just as bad relative to the competition? It may be half price, but so is the performance apparently! Are the AMD consumer card severely limited in compute on purpose? Or, as I suspect, is ROCm just holding the hardware back?

ROCm looks like a student project. Every year, they will get a few students to work on it, they do their assignments and go their merry way. There doesn't seem to be a vision and a will to make it a proper compute framework. Look at Intel and their level zero initiative, how fast that's moving, for comparison.

[aerospace]

How can a 5 year old 2080 Ti beats the current AMD consumer flagship?!

That is a good question. Do you have the opportunity to also test a 3070 or a 4070? We're buying RX cards mainly for the VRAM size...