Stable Diffusion Optimized for AMD RDNA2/RDNA3 GPUs (Beta) | Hacker News Reader

Stable Diffusion Optimized for AMD RDNA2/RDNA3 GPUs (Beta)

github.com

54 pointsirusensei15 comments

How faster this is from automatic1111's webui?

Also, Linux is supported right? Then why is this listed as a requirement

> AMD Software: Adrenalin Edition 22.11.1 for MLIR/IREE Driver Version 22.20.29.09 for Windows® 10 and Windows® 11 (Windows Driver Store Version 31.0.12029.9003)

Assume it's a binary blob that's needed, given it mentions Vulkan so they are skipping the ROCm method too.

Does rdna2/3 work with rocm drivers?

I've been using ROCm and HIP on a 6800 XT for some time. Don't know about rdna 3 though.

APUs don't.

And the 5000 series apus are vega cores, which are seemingly the amd tech that lends well to computing tasks…

There are Linux install instructions under the "Advanced Installation" expando.

A111 is a frontend for a bunch of things. What really determines performance here is Torch. If you are running it with an AMD GPU are probably Torch compiled with ROCm support which is a somewhat unpopular toolkit. This project seems to rely on Vulkan.

I see this is Windows only, but I don't fully understand what this is. The results are just images, and not a speedup. Has it not been possible to use AMD hardware for SD on windows before this?

>Windows Update may (depending how it's configured) automatically install a new official AMD driver that overwrites this IREE-specific driver.

This reminds me of when I used a Radeon 7870. The old catalyst control center drivers worked flawlessly but Windows kept installing crimson drivers which didn't work.

AMD really dropped the ball on this. Nobody even thinks of them or Intel for training or experimenting

It's easy to be dazzled by recent culminations of almost two decades of dominance in scientific computing, but don't forget how broad the field of vision and neural computing is.

I've been working with oneAPI and OpenVINO and using the new extensions for pytorch and tensorflow on an Arc GPU lately, and it's kind of amazing how fast things are moving.

Won't disagree they've slept on this but I'm pretty sure now that gpu compute has reached enthusiast hands people do think of AMD and whatever intel may bring to the table. Who knows? AMD tends to arrive late but with more available tech like FSR, Mantle, upstream kernel drivers etc.

Well, they are bad at advertising HIP support like the way NVIDIA does. But I'm on a 6900XT and I can run Pytorch nightly easy with it. No regrets.

pip3 install --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/rocm5.3

Kudos to Torch for achieving what Tensorflow did not want to do officially.

Worth mentioning that rocm packages are already available in debian testing.

apt install rocm-smi rocm-cmake rocminfo rocm-device-libs

I don't work with ML & co, so YMMV.

Reply on news.ycombinator.com