I have a question for the hackathon, specifically ...
# team-maintainers
f
I have a question for the hackathon, specifically related to GPU modules. What is our guidance these days on modules that run both on GPU and cpu. I know @Nico Trummer had some nice implementation, but I don't think it's officially documented somewhere (sorry if it is). I am asking because I am leading a GPU test + development group and would love to have some definitive guidance I can give people for how to implement GPU + CPU containers etc
No insights for GPUs? I am mostly looking at this proposal from @Nico Trummer and would love to formalize a bit more what nf-core recommends. As GPU tools become more popular and frequently used within nf-core, I think it's important we have some best practices for implementing it in modules and pipelines in place!
Ok, I have gotten zero feedback regarding GPU guidance, so I will fly this based on what @Nico Trummer has been doing so far during the hackathon and building upon this and after the hackathon adding some documentation and guidance about this to the docs. Speak up now or forever remain silent šŸ˜‰ šŸ˜›
n
I wrote a bit of documentation on the usage page of the #scdownstream docs too, that is less technical and more user-focused
f
Awesome, can you link that here please @Nico Trummer ?
Will mostly go by your implementations in #scdownstream to point people to, but I do think we need more high level guidance for GPU implementations for processes and pipelines.
p
Looks quite good. Some possible considerations (not sure if they're worth caring about now, but maybe easier now than later). • ROCm (AMD) is a thing, at least hardware-wise. It seems very likely to continue to be. Obviously, some softwares only support CUDA but what I hear from people more interested is that dual support codes aren't so rare • the
gpu
label doesn't give much guidance for what might be suitable models (mostly thinking memory consumption)
n
I can send the link tomorrow, remind me if I forget
šŸ‘ 1