Meauring data sent/received on the nvlink

Hi there,

I have a ubench to measure the data sent/received on the NVLINK. Will CUPTI based profiling work on a hopper-GPU system?

I’m aware I can use nvidia-smi to get nvlink stats. However, nvidia-smi seems to report numbers in KB which is somewhat too coarse grained for my application.

fyi, I based my ubench based on the nvlink bandwith sample found under the CUDA driver folder eg cuda-11.7/extras/CUPTI/samples/nvlink_bandwidth

Any advice would be great.

Thanks,
Chris

Unfortunately CUPTI doesn’t have NVLINK metrics profiling support for Hopper GPU. CUPTI sample nvlink_bandwidth is based on the Event API which is not supported on Turing and later GPU architectures. CUPTI Profiling APIs are supported on Volta and later GPU architectures including Hopper, but NVLINK metrics are not supported by these APIs.

Nsight Compute can be used to collect these metrics, check metrics with prefix nvlrx and nvltx. For ex - metric nvlrx__bytes provides total number of bytes received.

This topic was automatically closed after 9 days. New replies are no longer allowed.