This section outlines best practices and configuration examples for benchmarking GPU performance with CN5000. It includes recommendations for both NVIDIA and AMD GPUs with CN5000 tuning parameters to ensure optimal performance.

As of 12.1.1 software release, line-rate performance is achieved when using hfi1 BTS, and setting FI_OPX_HFISVC=1 at user runtime. On NVIDIA or AMD GPUs, you must also set FI_HMEM_CUDA_USE_DMABUF=1 or FI_HMEM_ROCR_USE_DMABUF=1, respectively.