how can you build a community software architecture (or community software, for that matter) on top of compilers and libraries that can't run anywhere but cray systems?

yep it's me 👋

Just throwing this out there: there are PRs right now for rolling ROCm and oneapi releases in Spack CI: https://github.com/spack/spack/pull/31781, https://github.com/spack/spack/pull/31450

Is there a plan for oneMKL to support CUDA (specifically cuSparse and cuFFT) and the corresponding ROCm libraries. Without this, oneMKL is restricted to Intel arch.

@Miroslav I think it was mentioned during the talk that oneMKL has a cuBLAS interface

such a thing exists -- we want you all to contribute 🙂

@Wileam Phan, cuBLAS does not cover the functionality in cuSparse and cuFFT, or anything in ROCm.

We are now building thousands of spack packages in CI on every PR: https://spack.io/spack-binary-packages/ -- this is a place where we would really like to see vendor compilers -- there are PRs for this for ROCm and oneapi above -- we may not get to all 6400 packages but we're trying

FWIW, there are key problems we solve that take ~4hrs -- so same-day turnaround. This was made possible w/GPUs

+1 to 2x on salary

it's also difficult for OSS projects to work with the NVIDIA developer license ... just saying

Fair

“here's a container with 6 things you can't distribute and 4 things you can -- good luck!"

thanks nvidia

I should probably say too, not having a clear license at all makes collaboration even more difficult, so please have a clear license and even better if it's one we can collaborate with easily

But since Mike specifically asked about OSS, I didn't want to say that at the time.

+100