warpx on H200 GPUs? #6083
-
|
Hi there, |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments
-
|
Hi @denisbertini, We run WarpX very often on A100 and some still on V100 GPUs, but have no direct access to a H100/H200 cluster right now. Generally speaking, the H200 is a refresh of the H100 with more memory bandwidth, so algorithms like a PIC code that are mostly memory-bound benefit from that increased BW.
Compiling WarpX for different Nvidia GPU generations is quite straight forward: You just need a new enough CUDA driver, CUDA toolkit (CUDA 12.0+ for SM 9.0), and need to set the SM during compilation of WarpX: export AMREX_CUDA_ARCH=9.0 # SM 9.0 for H200
cmake --fresh -S . -B build -DWarpX_COMPUTE=CUDA # ... more options ...
cmake --build buildIn practice, if you compile on the same node as you run, we even query the GPU with CMake and compile for that exact architecture. |
Beta Was this translation helpful? Give feedback.
-
|
Thanks for this info ! |
Beta Was this translation helpful? Give feedback.
Hi @denisbertini,
We run WarpX very often on A100 and some still on V100 GPUs, but have no direct access to a H100/H200 cluster right now.
Generally speaking, the H200 is a refresh of the H100 with more memory bandwidth, so algorithms like a PIC code that are mostly memory-bound benefit from that increased BW.
Compiling WarpX for different Nvidia GPU generations is quite straight forward: You just need a new enough CUDA driver, CUDA toolkit (CUDA 12.0+ for SM 9.0), and need to set the SM during compilation of WarpX: