I’m trying to get maximum HPL performance of GPU server equipped with 8 X SXM4 80GB A100 GPU and 2 socket AMD EPYC processor. I got the NVIDIA HPC-Benchmarks 21.4 image from NGC
When i ran HPL-AI(mixed precision) benchmark and checked nvlink usage, P2P communication between GPUs works well. But when i ran HPL (FP64) and checked nvlink usage, it seems the benchmark doesn’t support GPUs communication using nvlink. So i would like to know how to use nvlink P2P communication while running HPL benchmark test.
That version of HPL doesn’t use P2P communications, therefore you won’t see NVLink usage. In the future, things could change, of course. As someone who is just running codes (not writing them) you would have no control over whether an application is written to use P2P, or not. That version of HPL is not written to use P2P.