in the attached image, the blue line’s x axis is time. What is the y axis of the blue bar? In the image, it says “Avg CUDA kernel coverage” which is very unclear to me. It also says “Not SM occupancy”.
Also, what is the height of the red line:“avg number of memory operations”? Is it achieved throughput relative to the bandwidth?
There seems to be no mention of it in Nsight systems documentation as well.
This combined activity line on the GPU is a summary of the data. In order to try to help you see the difference between “there is some activity” and “the GPU is fully booked” without opening the GPU (down arrow on r-hand side) we have made the height of a given pixel width dependent on the average use underneath. It is approximate, so a full height bar is fully busy, and a ~25% busy GPU under that pixel is ~25% utilized.
You can refine the data and see more exact information (like whether the memory transfers are host->device, device->host, device->device) by clicking the arrow on the r-hand side.
You are correct that there should be an explanation of the summary bars (which can also be found in the CPU rows and the thread rows) in the documentation. I have added a feature request for this and assigned it to myself.
Thanks!
