[gbrown]

I’ve been playing with GPGPU on Julia lately also, and it really seems like things have come a long way in the last few years. Check out the Juliacon 2021 talk on GPU compute if you’re interested.

[e12e]

This one?

https://live.juliacon.org/talk/P8KJSW

[vchuravy]

Probably more this talk about CUDA 3.0 https://live.juliacon.org/talk/UGX8YR or the workshop https://www.youtube.com/watch?v=Hz9IMJuW5hU

[gbrown]

Yeah, I was thinking of the workshop.

[alhirzel]

Are AMD and NVidia GPUs on par nowadays? Or is it still an NVidia-first world when it comes to compute support.

[eigenspace]

There was actually a question about this earlier today: https://discourse.julialang.org/t/amdgpu-jl-status/71191

TLDR: The entire GPUArrays.jl test suite now passes with AMDGPU.jl. There are still some missing features and it is not as mature as the nvidia version, but this space is progressing rapidly, and benefited from the generic GPU compilation pipeline that was initially built for CUDA.jl

[snicker7]

Keep in mind that AMDGPU.jl requires ROCm, which is basically dead (no recent GPUs support it and none of those that do are consumer-grade).

The problem with AMD GPGPU is not software, it is that AMD literally does not care.

[jpsamaroo]

Definitely not dead; Vega is well supported, and with some tweaks, Polaris probably works too (although it definitely was broken in HIP around ROCm 4.0.0 or so).

I think AMD has some work to do on non-C++/Python ecosystem engagement for sure, but they've built a foundation that's quite easy to build upon and get excellent performance and functionality; AMDGPU.jl is a testament to that.

[JonChesterfield]

The gfx10 line (6800XT et al) probably work out of the box on a recent release. I think some are even officially supported. I test on a 5700XT which I don't think is officially supported. The change to 32 wide wavefronts took a while to resolve.

Rocm gets releases every few months or so. The llvm project part is mirrored to GitHub in real time.