Hacker News Reader

Comment by gymbeaux | Hacker News Reader

How do you have “some” CUDA support? I’m aware of the AMD HIP API that ports CUDA code to run on AMD GPUs, but that’s not CUDA at that point. Im also aware of the geohot (George Hotz) project for bringing native CUDA to AMD GPUs, but I think he abandoned it because AMD wasn’t throwing him any bones.

Cupy on Python is mostly what I use.

View on HN

This is what they are referring to: https://github.com/ROCm-Developer-Tools/HIPIFY

Pytorch can use the 'cuda' device on AMD GPUs, not talking about HIP. It works as a drop-in replacement if your ROCm API level matches your target CUDA level. (i.e matmul on fp16 won't be supported on old AMD GPUs for example)

This technically uses HIP. AFAIK, they ran the CUDA code for pytorch through HIPIFY

Fascinating, I didn't know about this. Thank you for sharing this bit!

I love this, but in practice I find things like this are yet another problem I have to work through. I have a hard enough time getting both the nvcc and nvidia-smi commands to work at the same time or two Nvidia GPUs (one internal and one external via TB3) to work on the same driver version. PopOS usually “just works”, most other distros it’s some kind of hassle, especially if the Nvidia GPU has to handle the display output too. I try to have an AMD APU as the CPU so I can have solid graphics support on Linux.

Anyway, do you know if this is just for pytorch? Is performance roughly 100% of what you’d get on an “equivalent” Nvidia GPU?

> but I think he abandoned it because AMD wasn’t throwing him any bones.

I haven't heard about the "bringing native CUDA to AMD GPUs", sounds really interesting. I did come across a picture of Geohot with a ton of AMD GPUs though, wasn't that enough for him or what?

HN is where I heard about it! About 3 months ago I think.

I think this is it: https://news.ycombinator.com/item?id=36189705

Reply on news.ycombinator.com