Running simulations on AMD Vega GPU

[amitch0852]

Hi,

I tried to create and simulate the model using CPU() function and it seems to be unresponsive. However, when I used the GPU function instead of CPU, it failed to un as my GPU has a different architecture than CUDA cores.
Please note that I am using Ryzen 5 3500 CPU with Vega 8 GPU. Is there any other way to simulate my model using VEGA GPU cores or any other architecture except CUDA?

Thank you.

[glwagner]

AMD GPUs are not currently supported by our API, as you have discovered, and the architecture "GPU" is a slight misnomer as it really means "nVidia GPU".

Note that Oceananigans may run slower on your Vega 8 GPU than your CPU. We usually use GPUs specifically designed for scientific applications like the Titan V or V100 for research runs. I believe CPUs tend to be competitive or faster than mid-range GPUs defined for gaming and laptop / desktop use. (Though I should say that the beefiest gaming GPUs like the Quadro P6000 do deliver speed ups.) @ali-ramadhan might know more about this.

For future reference, support for AMD GPUs may not be far away (especially after #1567) as there is some support in the underlying KernelAbstractions package. There's a few pieces of code that we would have to generalize to get this work (mostly figuring out the API for specifying an "AMDGPU", generalizing arch_array, and squashing stray usages of CuArray that surely exist). I'm not sure there have been any honest efforts in this direction. There is some discussion of using AMD GPUs with KernelAbstractions here: JuliaGPU/KernelAbstractions.jl#235.

Oceananigans does work on CPUs, though. If you share your code, your julia version, and the julia environment you are running your code in, perhaps we can help. Compile time has increased recently, especially using julia 1.6.