Support to NVIDIA Grace-Hopper Systems

[ccchen057]

Is there any plan to support or test on NVIDIA Grace-Hopper systems (Arm Architecture and NVIDIA H100)?

[jcandy]

@sfiligoi Any thoughts?

[sfiligoi]

It should just work.
There is nothing special about Grace Hopper... just another "NVIDIA-GPU-based system"
All you need is access to the HW and a recent compiler (ideally, NVIDIA HPC SDK)

[sfiligoi]

The only potential "problem" we may encounter is a performance one;
since the CPU and GPU can access each other memories, OpenACC/OpenMP may have a slightly different semantics.

e.g. We currently assume we can initialize a buffer on the CPU and then copy that buffer to the GPU, and get optimal performance.
If the buffer would instead reside wherever it was "first touched", that may result in expensive remote data accesses if we are not careful.

Again, not a showstopper, but something to consider if we get significant G-H compute resources in the future.

[lorenzozanisi]

Hello, have you had any success or issues running CGYRO on GH200/GH100 chips in the end?

[sfiligoi]

@lorenzozanisi Unfortunately, I don't have access to such a system, so no.
If you have access to one, I would be happy to help with any issues you encounter.