feature request: a feature that tells you the average GPU utilization over a period of time. Currently, the skyblue bar in the “CUDA” row indicates at what time the GPU was used. A feature to get the averge of: time used by GPU / total scrolled time, would be helpful/
question on Nsight Compute and nvidia-smi: I know that nvidia-smi’s utilization is the time that GPU did some computation ( as explained very well by Crovella here: https://stackoverflow.com/questions/40937894/nvidia-smi-volatile-gpu-utilization-explanation/40938696#40938696 ). so is it measuring the same thing as the sky-blue line in the “CUDA” row is measuring, or is it taking other things – potentially the time spent by GPU to call CUDA API – into the “utilization” metric as well?
Currently, the skyblue bar in the “CUDA” row indicates at what time the GPU was used.
You are referring to the UI from which tool?
Nsight Compute UI tool
I suppose you are referring to the timeline view in Nsight Systems.
Oh yes. Sorry for my confusion. I meant to say systems not compute.
I moved this to Nsight Systems - assuming that you are on x86 Linux platform.
Which platform are you using Nsight Systems on?
Target is CentOS7 and host is Mac Nsight Systems UI.
With the most recent shipped version of Nsight Systems, 3.1.85, shipped last Thursday, we now have a beta exporter that allows you to export the data into SQLite and search it as you like. This will give you the information you are looking for.
The documentation for this feature isn’t fully integrated, but you will find it in the installation directory
/Documentation/nsys-exporter. It contains directions for how to use it, and some recipes for getting information.
In a future version, later this year, we will have some common statistics available from a simple script and instructions about how to change that script to pull out anything you specifically need.