Lisa Su Reaffirms Commitment To Improving AMD ROCm Support, Engaging The Community

[Linuxhippy]

As far back as I can remember, AMD always had problems with the software-support of their products - because AMD views itself as a hardware company and software for them is only a necessity (=causes cost) to sell their hardware. This is also the reason why most of the propretary driver developers are located in cheap development countries - AMD is not willed to spend money on developing software. The real exception seem to be the open-source driver developers that joined over the last view years,.

This is the reason why Nvidia got so big in AI with their CUDA effort - they understood very clearly that to a good deal the written software defines which hardware will be sold.
They understood you have to invest and spend a lot of money in software and did a good job with their APIs. Suddenly a lot of software appeared which was vendor-locked to nvidia. You can change GPU microarchitecture with a new driver, but Cuda software is here to stay and requires - you name it - Nvidia hardware basically forever.

Whenever AMD announces something, it is only half-hearthed and again - viewed as a necessity to compete with nvidia. It is always announced with a lot fanfare, but after the initial investment and without immediate success things are not pusured as motivated as necessary to make a difference...

[Developer12]

Yeah right!
When will it be fully open source and installed by default like Mesa?

It is fully open source, and there's a baseline ROCm driver upstream. The biggest issue is that AMD themselves only support a small subset of cards.

It turns out AMD also kinda screwed themselves here, because compiled HIP code is way less portable between cards than CUDA PTX is.

To quote wikipedia:

Parallel Thread Execution (PTX or NVPTX) is a low-level parallel thread execution virtual machine and instruction set architecture used in Nvidia's CUDA programming environment. The NVCC compiler translates code written in CUDA, a C++-like language, into PTX instructions (an assembly language represented as ASCII text), and the graphics driver contains a compiler which translates the PTX instructions into the executable binary code which can be run on the processing cores of Nvidia GPUs.

Afaik HIP code has to be compiled for like, a specific card or family of cards. Meanwhile, OpenCL 3.0 like with rusticl is kinda the ideal solution, because it means compiling into SPIR-V or whatever which is portable across ALL GPUs. Not that a GPU vendor would like it if you did that.

[pegasus]

Also largest hpc systems are now making use of Rocm stack ... And it anyone, they will get it in shape. It's just a question of time.

[brucethemoose]

It turns out AMD also kinda screwed themselves here, because compiled HIP code is way less portable between cards than CUDA PTX is.

To quote wikipedia:...


Afaik HIP code has to be compiled for like, a specific card or family of cards. Meanwhile, OpenCL 3.0 like with rusticl is kinda the ideal solution, because it means compiling into SPIR-V or whatever which is portable across ALL GPUs. Not that a GPU vendor would like it if you did that.

AFAIK its more complicated.

CUDA can actually target specific generations as architectures (with sm_75 being turing, for instance), and supporting multiple architectures is why some CUDA (like PyTorch) are hilariously large. Other apps can compile code on the fly or target architectures natively.

Meanwhile, I think ROCM compiles more code on the fly? The rocm pytorch releases are significantly smaller for whatever reason, but I am way out of my depth here.

[finalzone]

$ sudo dnf install rocm*

does not work on Fedora38... blender cycles does not work.

rocm-hip-runtime is currently unavailable in the main Fedora repository. However, you can use the amd driver to get hip-runtime available on https://www.reddit.com/r/linuxquesti..._on_fedora_37/

[bearoso]

I typically dislike attacking character, but I've seen the Hotz meltdown on Youtube and it's the most childish thing I've seen in a long while.
Screaming for 10 good minutes and announcing that he was dropping AMD for good because the answer of his AMD contact was unacceptable to him...

Many people thought he had talent with all the PS3 hacking and whatnot, but when he joined Twitter, just like Musk, he revealed himself to be an arrogant man-child.

Then, because of said publicity, further scrutiny revealed he stole his PS3 and iOS exploits from other people, which means he lacks talent as well.

So just ignore him.

[sophisticles]

As I have said before, just buy an Nvidia card and call it a day.

I can buy a used Quadro for under $100 that will provide more than enough compute performance for most people and i won't have any driver issues at all.

[Quackdoc]

For who?
You and I can't install reaffirmed commitments. It's a mostly usage cry: "rats, please don't leave this boat, we swear we'll plug the holes in it".

In the words of master Yoda: "Do! Or do not!"
"Reaffirming commitment" is nothing in this context.

hotz is pretty reputable, if he says it's going to happen and he has faith that it will, then I would trust him on it, also shortly after the rant and him posting that he was done, AMD posted this job which is pretty much explicitly what he was asking for https://ca.linkedin.com/jobs/view/sy...amd-3602069567

I do actually hope that things will start changing for the better now

[timofonic]

ROCm is trash, RustiCL and full MESA integration is the way forward.

They need to make it not only better than CUDA in all means in an Open Source way, but make it really work on all their hardware even consumer grade ones.The rest is just astroturfing.

We are tired of empty words, Lisa Su and rest of AMD/Radeon Team...

With this crap, AMD is making the Nvidia hegemony in GPGPU even bigger. I give them "thanks" in a very sarcastic way

Software is very important, AMD. Hire good high-end system developers and throw the crappy code. Intel is a lot better at software than AMD and Open Source friendly, that made them stay alive despite their crappy hardware.

[Eirikr1848]

I’m an optimist but: for certified workloads simply using Mesa isn’t the best path forward.

I mean: could the open source stack be “upstream” and provide OpenGL, Vulkan, and OpenCL? Sure, just as it is now.

Then what? Downstream: AMD patches and certifies versions X often, building and expanding support to including everything from Southern Islands and Sea Islands consumer, pro and compute products through modern?

HIP/HSA then operates on top of this stack, providing cross-platform compute and CUDA/PTX translation capabilities across everything from Kabini to the 2x CUs in Ryzen 7000 to to the R9 Fury and latest Instinct cards?

I mean - how would this ACTUALLY look though, compared to how it looks now - outside of even higher quality open-source drivers?

(that being said… in the case of RADV vs AMDVLK: RADV wins in most cases for performance… so is having AMD committed to Mesa solutions… really the best path?)

ROCm is trash, RustiCL and full MESA integration is the way forward.

They need to make it not only better than CUDA in all means in an Open Source way, but make it really work on all their hardware even consumer grade ones.The rest is just astroturfing.

We are tired of empty words, Lisa Su and rest of AMD/Radeon Team...

With this crap, AMD is making the Nvidia hegemony in GPGPU even bigger. I give them "thanks" in a very sarcastic way

Software is very important, AMD. Hire good high-end system developers and throw the crappy code. Intel is a lot better at software than AMD and Open Source friendly, that made them stay alive despite their crappy hardware.