Nsys Does not Track CUDA Api events

Used the command below, but no cuda events were captured(no kernels, no memcpy or memset).

nsys profile -w true --trace=cuda,nvtx,cudnn,cublas -s cpu --force-overwrite true --cudabacktrace=true -x true -o my_profile ./devel/lib/lidar_centerpoint/lidar_centerpoint_node

  • NVIDIA Nsight Systems version 2020.3.4.32-52657a0
  • Driver 515.65.01
  • CUDA 11.1

That screenshot is only showing activity on the CPU side of the system. Can you show me what you see on the GPU side?

That is a super old version of both the CUDA ToolKit and of Nsys. Would you be willing to try with a newer version?

Unless you really need them, I would recommend not collecting --cudabacktrace, it is a very high overhead trace funtion.

How can I upgrade Nsys and toolkit without changing my CUDA version?

Nsight Systems is available separately for download via NVIDIA Nsight Systems | NVIDIA Developer

If you get it from there you will also get the 2022.5 version which just shipped this week.