I have a holoscan application written in python which uses CuPy or Numpy(depending on user choice) running on my notebook. When using CuPy, both nvTop and Nsight show GPU activity which is as expected, however, when using Numpy, nvTop shows about 15% GPU activity while there is no GPU activity in Nsight.
Why?
Is there any difference in what these two tools actually record?
Any resources to shed some more light on this would be of much help.
Thanks!
EDIT:-
The nvtop output shows two processes for the holoscan application one with the type graphic and another with compute.
Please provide more details of the Nsight Systems session you captured. Which trace options did you enable? Did you capture GPU Metrics? Can you share the .nsys-rep file?
Hi,
I enabled trace options for cuda, nvtx and osrt and captured gpu metrics.
Unfortunately, my code-base is closed-source, and the .nsys.rep file contains internal function names and data that I’m not able to share publicly.
Is there something else which i can share and might help?
So my implementation uses some video rendering which is getting recorded in nvtop output as a GPU activity.
Even though my previous trace was including GPU metrics, it didn’t include the –gpu-video-device=0(0 being my GPUs ID) after which the GPU activity recorded by nvtop was also recorded by nsight systems.