I have not found this function when using compute tool, I want to know whether/when those hardware engines are active/inactive, for example tensor cores and cuda cores in SM, is this idea possible?
Hi, spring_wind
Please see if “System trace” activity can meet your requirement.
Why this function is disabled in my envrionment, :(
It is weird. It should not be disabled. Can you get a latest Nsight Compute version to check ? NVIDIA® Nsight™ Compute 2023.3 is now available
Application: NVIDIA Nsight Compute
Version: 2023.3.0.0 (build 33266684) (public-release)
Yes，I use the latest version as above. But I use ssh connect way to profile my kernel, it that the reason?
Yes. That’s the reason. We don’t support remote trace in Nsight Compute now.
I read this introdution about the “System Trace”, but it is not what I want to get.
Actually, what I want to get is all compute engines(including cuda cores and tensor cores)'s activity in timeline withine a kernel, interaction between CPU and GPU is not included.
Thanks for confirmation. I will check with our dev and reply you back soon.
have you get any feedback from your dev?
Hi, @spring_wind
Sorry for the late response ! A new feature of PM Sampling in Nsight Compute (see https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#pm-sampling for more details) was first introduced with version 2023.3.0. This allows users to collect metrics over time during the kernel execution. This will support also the metrics you requested in the comment from Nov 9.