Run CUDA, unmodified, on AMD GPUs

Re:Run CUDA, unmodified, on AMD GPUs by mappu on Monday July 15, 2024 @08:57PM (#40972822)

AMD's definition of "support" I think is different than what people expect, and pretty misleading - ROCm itself will run on almost anything, back as far as the RX 400/500 series:

https://en.wikipedia.org/wiki/ROCm#:~:text=GCN%205%20%2D%20V...

Stable Diffusion ran fine for me on RX 570 and RX 6600XT with nothing but distro packages.

[ Parent | on HN ]

Re:Run CUDA, unmodified, on AMD GPUs by imtringued on Tuesday July 16, 2024 @04:18AM (#40974541)

I don't buy it. Even running things like llama.cpp on my RX 570 via Vulkan crashes the entire system.

[ Parent | on HN ]

Re:Run CUDA, unmodified, on AMD GPUs by Nab443 on Tuesday July 16, 2024 @04:59AM (#40974728)

The last time I checked, I was stuck with a pretty old kernel if I wanted to have the last version of ROCm available for my rx470. It's compatible at some point in time, but not kept compatible with recent kernels.

[ Parent | on HN ]

Re:Run CUDA, unmodified, on AMD GPUs by mappu on Tuesday July 16, 2024 @06:04PM (#40980763)

It's the responsibility of your distro to ship things that work together,

[ Parent | on HN ]

Re:Run CUDA, unmodified, on AMD GPUs by slavik81 on Tuesday July 16, 2024 @02:00PM (#40978762)

There are out-of-bounds writes in the BLAS libraries for gfx803 GPUs (such as the RX 570). That hardware might work fine for your use case, but there's a lot of failures in the test suites.

I agree that the official support list is very conservative, but I wouldn't recommend pre-Vega GPUs for use with ROCm. Stick to gfx900 and newer, if you can.

[ Parent | on HN ]