AMD ROCm Going Open-Source: Will Include Software and Hardware Documentation

https://wccftech.com/amd-rocm-going-open-source-will-include-software-stack-hardware-documentation/

by ksec • 2 years ago

86 33 2 years ago

wccftech.com

otoburb• 2 years ago

From the AMD Radeon tweet referenced in the article[1] (emphasis mine):

"Open sourcing additional portions of our software stack and more hardware documentation."

So, not the whole driver stack, but still a positive step forward in the right direction.

[1] https://x.com/amdradeon/status/1775261152987271614

Fubarberry• 2 years ago

ROCm has lagged behind CUDA, which is definitely unfortunate for linux users since AMD graphics cards have less issues than the nvidia ones.

Hopefully this will let ROCm be a valid competitor.

hkwerf2• 2 years ago

I do love AMD because its drivers are open source as opposed to nVidia. However, is "less issues" really the case? I sure hope not, for nVidia's sake.

My AMD graphics experience, on APUs and dedicated GPUs, has been plagued with basic issues and random crashes. AMD cards and their driver/firmware issues are without a question the IT product that affects my daily life most.

orbifold2• 2 years ago

I tried AMD exactly once for Linux graphics, it was so unstable that I bought a NVIDIA card within a month and have never looked back. Potentially NVidias approach of staying as far away as possible from the Linux kernel abstractions is much saner than to play ball with them.

noch2• 2 years ago

I do love AMD because its drivers are open source as opposed to nVidia.

AMD's drivers are not really more open that Nvidia's. Similar to Nvidia's Open GPU Kernel Module's[0], AMD's opensource drivers are mostly a shim that wrap firmware blobs[1] in which the functionality you really care about is contained.

[0] https://github.com/NVIDIA/open-gpu-kernel-modules/discussion...

[1] https://github.com/geohot/7900xtx/

maeln• 2 years ago

From what I remember from the ATI days (aah, sweet memories), it was often the case of either you card was working great and the driver was much more up to date with the latest kernel tech, or the card was a mess to make work under linux. With NVidia, you were more sure that your car was going to work, but they were often not implementing certain kernel tech, or they had their own weird way to do things, which often mean that you had to configure Xorg and some app with some weird workaround for NVidia.

It seems like it didn't change much since those times. AMD card are still a fifty fifty, and NVidia still make open source maintainer write documentation for their card ( https://wiki.hyprland.org/Getting-Started/Master-Tutorial/#n... )

asksomeoneelse• 2 years ago

In my experience, AMD still has some massive stability issues with the VBIOS of their GPUs. But it's not a component that the users can officially update on their own (whichever version the graphic card comes with when you receive it, is what you are stuck with), and most users are not even aware of its existence (even on Windows, the official driver doesn't provide any way to update it, or even check its version). Resulting in two seemingly similar cards, with the same drivers and firmware, behaving wildly differently in some case depending on the production batches the card was part of. Ask me how I know.

smallmancontrov2• 2 years ago

AMD graphics cards have less issues

Not on compute. That's the real problem. Compatibility would have happened a decade sooner if AMD's compute drivers actually worked.

meragrin_• 2 years ago

As far as I am aware, ROCm itself has been open since its release in 2016. Whatever they were trying to convey will have little impact on ROCm being a valid competitor. Frankly, AMD should just trash ROCm and go all in on Intel's oneAPI. AMD seems helpless on software and reports are Google and Qualcomm have interest in pushing oneAPI.

nisegami• 2 years ago

I wonder if this is what George Hotz was alluding to when he announced tiny corp would be reversing-reversing course and offering AMD boxes as an option.

noch2• 2 years ago

No. AMD isn't going to do the right thing and will not open source firmware. What a wasted opportunity by Team Red.

https://twitter.com/__tinygrad__/status/1775264921468764205

cycomanic2• 2 years ago

I'm all for demanding as much open source as possible, but I'm not aware of any big company releasing the source for their firmware. Intel who has often been hailed as the paragon of open source is not releasing the firmware of their ( less capable) cards, nor are they releasing the microcode to their CPUs (neither is AMD).

Open source drivers with closed source Firmware is still much better than the closed source mess we have with Nvidia

sylware2• 2 years ago

Hardware programming manual incomplete? We would not have mesa neither the linux kernel modules without a complete with erratas hardware programming manual.

hackernudes• 2 years ago

AMD employs people to write the open source drivers. They have access to internal resources. I think it is possible that the public docs are incomplete.

sylware2• 2 years ago

Code does not replace the hardware programming manual with up-to-date erratas. If this is the case, hopefully AMD is not hypocrit on that matter, namely releasing on purpose outdated or incomplete hardware programming manual/erratas to force people to use their code (which could be very very bad, look at that c++ abomination which is libaddr, which should be plain and simple C99).

instagraham• 2 years ago

Is AMD's situation even fixable? Since 2022, if you were doing AI, was there any reason at all to have an AMD GPU over an NVIDIA? If you're just someone who wants to try out some of the cool projects shared on HN alone, you'd have been left out of the conversation entirely for the last 2 years.

The worse part is that they had an internal project that could've enabled CUDA applications to run with ROCm, but they stopped funding it [1]

This is all to say - AMD GPUs make sense if you're a gamer and if they're cheaper than an NVIDIA where-ever you buy your GPUs. But otherwise, it feels like buying a SymbianOS phone while Android and iOS were into their second or third generations.

1 - https://www.xda-developers.com/nvidia-cuda-amd-zluda/#:%7E:t...

m00x• 2 years ago

Running CUDA applications on AMD was always going to be slower. They have slightly different architectures, and CUDA is completely closed source. They would be committing to always being a step behind.

If you want to use both, you can write SYCL.

bryanlarsen• 2 years ago

The fraction of people who are doing AI at the CUDA level is far smaller than the fraction of people who are doing it at the TensorFlow or Ollama levels.

brokencode• 2 years ago

Even if AMD could allow CUDA to run on Radeon GPUs from a technical perspective, how would that work out legally?

It was my understanding that a big part of what makes CUDA so good is all the math libraries that Nvidia wrote and optimized for their GPUs. Are those libraries licensed to run on non-Nvidia GPUs?

To me, it seems like what AMD needs to do is invest heavily in the ROCm ecosystem to bring it up to the level of CUDA. But yeah, it probably would take a long time to catch up.

You have to hand it to Nvidia here. They invested in the GPGPU vision long before it was economically viable. Their long term focus has allowed them to reap incredible profits from AI now and crypto before it.

Contrast this with the Intel approach of halfheartedly trying a new market every couple years and abandoning it after it’s not immediately successful. See Larrabee for the relevant example. But also Optane and their 5G cell modem as other notables.

nixgeek• 2 years ago

Looks like geohot is now advertising Team Green and Team Red options with the latter being $10k cheaper.

tinybox green - 6x NVIDIA RTX 4090 ($25k)

You came here to do ML and you want it to just work. You like checking out the latest open source ML stuff on GitHub. gpt-fast. You want your Attention to be CUDA quality FlashAttention. You want FP8. A solid box for the ML researcher. A solid box for the tinygrad developer. A solid box for everyone. It's just a little expensive.

tinybox red - 6x AMD 7900XTX ($15k)

You came here for the hardware.

You want to be a part of the journey. Make AMD great, with or without their help. Sure, it works for ML, mostly. Just maybe you have to reboot it sometimes. tinygrad does work well enough to get AMD on MLPerf at least. You buy the red box because you want to mess with kernel drivers and shader engines. You know what a warp is. You like that you can see the MMIO with umr. RDNA3 is documented! You want to get every bit of perf out of the metal.

The specs are similar to green, except it's $10k cheaper.

smallmancontrov2• 2 years ago

These are wise descriptions of the NVIDIA/AMD competitive landscape.

roenxi• 2 years ago

I don't know what AMD is planning here; but tweets and a bug tracker is not useful progress. I am not optimistic that this will be meaningful.

That being said, if they start open sourcing the firmware that would potentially be huge. And if they do actually fix some of geohot's bugs that would potentially be significant. There seems to be a small and finite set of bugs on any given AMD card that are blockers. If, by hook or crook, they fix that list of bugs then the card will probably be a lot more useful for machine learning. If they stop with the stupid crashes, everything else seems to be ok. They only need to do BLAS and data transfers reliably.

But knee-jerk bugfixing in response to public pressure is a really bad sign. This still wouldn't be the thoughtful root cause analysis that leads to high quality software. Plan->Do->Check->Act; not Panic->Tweet->Code->Crash.

AMD ROCm Going Open-Source: Will Include Software and Hardware Documentation

Related Stories