Oceananigans.jl
Oceananigans.jl copied to clipboard
CliMA
Oceananigans on AMD GPUs
It looks like KernelAbstractions.jl has added experimental support for AMD: https://github.com/JuliaGPU/KernelAbstractions.jl/releases/tag/v0.6.0
Should we try to make the code more GPU-architecture-agnostic based on that? From what I've heard, end-user level AMD GPUs are generally cheaper than Nvidia ones for similar performance. So running Oceananigans on AMD could potentially make it significantly easier for people to run simulations on their personal computer instead of a cluster.
Thanks for sharing this @tomchor .
Do I take that to mean that if we try and run Oceananigan on an AMD GPU then very likely the code will fail because KerelAbstractions.jl (and possibily others) will fail?
It would be great to be able to run the library on any GPU but I guess a lot of this depends on KerenAbstractions, and other libraries. Do people know what other libraries we depend on that would not necessary run on AMD GPUs?
I don't know a lot about this (@ali-ramadhan and @glwagner can give you a much clearer picture) but I think up until last week KernelAbstractions.jl straight up didn't support AMD. It relied on defining CUDAKernels which are specific to Nvidia GPUs. Now I believe they also added ROCKernels, that are meant to take care of the AMD case.
However I think we still need some work on our end since Oceananigans imports CUDA and uses some CUDA-specific functions in lots of places. I think something needs to be done in these instances.
Agree that AMD may indeed be the future and we should take a look at this soon.
