What I wrote about this on twitter: https://x.com/HotAisle/status/1914549886185611627
https://open-iov.org/index.php/GPU_Support#AMD
https://github.com/GPUOpen-LibrariesAndSDKs/MxGPU-Virtualiza...
As "AI" use cases mature, NPU/AIE-ML virtualization will also be needed.
1) Support your graphics cards on linux using kernel drivers that you upstream. All of them. Not just a handful - all the ones you sell from say 18 months ago till today.
2) Make GPU acceleration actually work out of the box for pytorch and tensorflow. Not some special fork, patched version that you “maintain” on your website, the tip of the main branch for both of those libraries should just compile out of the box and give people gpu-accelerated ML.
This is table stakes but it blows my mind that they keep making press releases and promises like this that things are on the roadmap without doing thing one and unfucking the basic dev experience so people can actually use their GPUs for real work.
How it actually is: 1) Some cards work with rocm, some cards work with one of the other variations of BS libraries they have come up with over the years. Some cards work with amdgpu but many only work with proprietary kernel drivers which means if you don’t use precisely one of the distributions and kernel versions that they maintain you are sool.
2) Nothing whatsoever builds out of the box and when you get it to build almost nothing runs gpu accelerated. For me, pytorch requires a special downgrade, a python downgrade and a switch to a fork that AMD supposedly maintain although it doesn’t compile for me and when I managed to beat it into a shape where it compiled it wouldn’t run GPU accelerated even though games use the GPU just fine. I have a GPU that is supposedly current, so they are actively selling it, but can I use it? Can I bollocks. Ollama won’t talk to my GPU even though it supposedly works with ROCm. It only works with ROCm with some graphics cards. Tensorflow similar story when I last tried it although admittedly I didn’t try as hard as pytorch.
Just make your shit work so that people can use it. It really shouldn’t be that hard. The dev experience with NVidia is a million times better.
Same experience here. Installing ROCm and Ollama on my box were both dead simple and everything worked right out of the box. Using an RX 7900 XTX card, FWIW.
Because, there are tons of IP and trade secrets involved in driver development and optimization. Sometimes game related, sometimes for patching a rogue application which developers can't or don't fix, etc. etc.
GPU drivers are ought to be easy, but in reality, they are not. The open source drivers are "vanilla" drivers without all these case-dependent patching and optimization. Actually, they really work well out of the box for normal desktop applications. I don't think there are any cards which do (or will) not work with the open source kernel drivers as long as you use a sufficiently recent version.
...and you mention ROCm.
I'm not sure how ROCm's intellectual underpinnings are but, claiming lack of effort is a bit unfair to AMD. Yes, software was never their strong suit, but they're way better when compared to 20 years earlier. They have a proper open source driver which works, and a whole fleet of open source ROCm packages, which is very rigorously CI/CD tested by their maintainers now.
Do not forget that some of the world's most powerful supercomputers run on Instinct cards, and AMD is getting tons of experience from these big players. If you think the underpinnings of GPGPU libraries are easy, I can only say that the reality is very different. The simple things people do with PyTorch and other very high level libraries pull enormous tricks under the hood, and you're really pushing the boundaries of the hardware performance and capability-wise.
NVIDIA is not selling a tray full of switches and GPUs and require OEMs to integrate it as-is for no reason. On the other hand, the same NVIDIA acts very slowly to enable an open source ecosystem.
So, yes, AMD is not in an ideal position right now, but calling them incompetent doesn't help either.
P.S.: The company which fought for a completely open source HDMI 2.1 capable display driver is AMD, not NVIDIA.
I'm aware that 8GB Vram are not enough for most such workloads. But no support at all? That's ridiculous. Let me use the card and fall back to system memory for all I care.
Nvidia, as much as I hate their usually awfully insufficient linux support, has no such restrictions for any of their modern cards, as far as I'm aware.
And given the tooling they are adding full steam ahead for Python, I even wonder if Mojo will manage to get enough mindshare, let alone what AMD and Intel are not doing.
All the stuff works even if it's not officially supported. It's not that hard to set a single environment variable (HSA_OVERRIDE_GFX_VERSION).
Like literally, everything works, from Vega 56/64 to ryzen 99xx iGPUs.
Also, try nixos. Everything literally works with a single config entry after recent merge of rocm 6.3. I successfully run a zoo of various radeons of different generations.
And no, just because the three random cards you have work doesn't mean "everything works". Just tried an MI300A a few months ago... I just wanted to test ollama as this is one of the hottest applications for GPU acceleration now, it will surely be well supported right? First, the gfx version listed for it in the ollama docs is wrong - but OK, figured it out. Then Tried some random models with it, the only output it ever generates is GGGGGGGGGGGGG. Apparently only fp16 models work, nothing more quantized. So I pick one explicitly. Then it's slower than running on the cpu in the same system.
Thanks but no thanks; this cost me two days when Nvidia just works first try.
And they open sourced their GPU driver, which is a massive plus in my book.
If Radeon got virtualization support, it would make it the perfect GPU for running video games (or other GPU applications) in virtual machines - you wouldn't need to mess with PCIe passthrough anymore.
ATI started as a much more closed company. Then, they pivoted and started to open their parts. They were hamstrung at the HDCP at one point, and they decided to decouple HDCP block from video accelerators at silicon level to allow open source drivers to access video hardware without leaking/disabling HDCP support. So, they devoted to open what they have, but when you have tons of legacy IP, things doesn't go from 0 to 100 in a day. I want to remind that "game dependent driver optimization" started pre 2000s. This is how rooted these codebases are.
NVIDIA took a different approach, which was being indifferent on the surface, but their hardware became a bit more hostile at every turn towards nouveau. Then they released some specs to allow "hardware enablement" by nouveau, so closed source drivers can be installed, and the card didn't blank-screened at boot.
Then, as they fought with kernel developers, with some hard prodding by Kernel guys and some coercing by RedHat, NVIDIA accepted to completely remove "shim" shenanigans, and moved closed bits of the kernel module to card firmware by revising card architecture. It's important to keep in mind that NVIDIA's open drivers means "an open, bare bones kernel module, a full fledged and signed proprietary firmware which can be used by closed source drivers only and a closed source GLX stack", where in AMD this means "An open source kernel module, standard MESA libraries, and a closed source firmware available to all drivers".
It was in talks with nouveau guys to allow them to use the full powered firmware with clock and power management support, but I don't know where it went.
The CUDA environment is also the same. Yes it works very well, and it's a vast garden, but it's walled and protected by electrified fence and turrets. You're all in, or all out.
Probably because they don't have an open-source driver for linux and they can focus on the proprietary one.
https://wiki.archlinux.org/title/Ollama 3 commands to install and run an LLM.
Because it's like 2-4 times cheaper than to go nvidia?..
> the three random cards you have
It's more than 3 random cards. I run 6900 XT, 7900 XTX, W7900 Pro, VII, VII Pro, Vega 56, Vega 64, 6800 XT, 5700 XT plus I've experimented with a 9950 iGPU, a 5xxx series iGPU and the only thing which didn't work was 3400g iGPU.
> Apparently only fp16 models work
fp8 works for me
That still makes nvidia way cheaper compared to the time wasted getting that shit to run plus debug an unknown amount of issues down the road. I told my boss "look I wasted two days on this and it's barely working, I could've done actual work here". Two other colleagues poked at it some more without success. going Nvidia was a no-brainer then.
If you decide to dispose of that MI300A of yours, send it to me, I'll pay for shipping and even buy you a pizza.
Try it out! Benchmarks here: https://github.com/ggml-org/llama.cpp/discussions/10879
(ollama doesn’t support vulkan for some weird reason. I guess they never pulled the code from llama.cpp)
ROCm is a train wreck to compile from source but can be done with sufficient bloodymindedness.
The RX6600 is a gfx1032. I used a gfx1010 for ages with this stuff. Seems likely it'll run for you if you ignore the "supported cards" list, which really should be renamed to something that antagonises people less.
> ROCm is a train wreck to compile from source but can be done with sufficient bloodymindedness.
Yeah, I did that myself. Not impossible, just a bit annoying and time consuming. The issue I ran into then was exactly picking the gpu model (incredible that this is even necessary) and not having the gfx1032 available, see https://github.com/void-linux/void-packages/issues/26415#iss... for what I was following back then. I tried to edit the configuration for the gfx1032 anyway, but it did not succeed.
Side note: Already having to know which card corresponds to which code is annoying, and completely unnecessary. They could also just map the consumer facing name. But that would be too easy I assume.
> Nvidia, as much as I hate their usually awfully insufficient linux support, has no such restrictions for any of their modern cards, as far as I'm aware.
In fact, I regularly run llamafile (and sometimes ollama) on an nvidia dGPU in a laptop, with 4GB of VRAM, and it works fine (ish... I mostly do the thing where some layers are on the GPU and some are CPU; it's still faster than pure CPU so whatever).
They can't open source their proprietary drivers even if they wanted to because they don't own all of the IP and their code is full of NDA'd trade secrets. AMD isn't paying two different teams to do the same work because they like wasting money.
On the other hand, the card enablement sequences are open for AMD and Intel. AMD only protects card's thermal and fan configuration data to preventing card damage, AFAIK. You can clock the card and use its power management features the way you like. For NVIDIA, even they are out of reach.
AMD's open drivers work way better than NVIDIA's closed ones, too. I have never seen how a single application refused to launch until I used NVIDIA closed drivers.
Yet, HDMI forum said that they can't implement an HDMI2.1 capable driver in the open, with some nasty legal letters.
I have a couple of friends who wrote 3D engines from scratch and debugged graphics drivers for their engines for a living. It's a completely different jungle filled with completely different beasts.
I think being able to call glxinfo on an AMD card running with completely open drivers and being able to see extensions from NVIDIA, AMD, SGI, IBM and others is a big win already.
What does the 3d engine look like? Custom made AAA in companies such as EA or Ubisoft?