GPU usage - who do I believe?

Running the Textile demo I was puzzling over the lumpiness of the GPU usage as shown in the app and thinking, gee, wouldn’t it be so much faster if the GPU was being used more…

And then I looked at the Task Manager CUDA usage (and Copy too, which is a good indicator of workload changing).

I can’t reconcile the two views: why does the PL GPU graph peak so strongly and otherwise indicate zero usage when task manager shows the GPU as fluctuating much more and probably having a higher average…?

(NB I also couldn’t reconcile the CPU usage of ~20% with other measures, but the discrepancy is less dramatic)

Hi @JulianSMoore,
For complete transparency, we are using GPUtil to fetch your GPUs and then we use their .load property to see how much it’s being used.
This is then being averaged over all your GPUs at the moment (later on we will show for individual GPUs).
Maybe the issue is in the averaging, if you have multiple GPUs?

Depends what you mean by “GPU” :wink: There’s only 1 CUDA capable (nVidia) GPU, but the motherboard has Intel 630 UHD graphics GPU.

And the flat part of the curve seems (to me) to be too low to be 50% of the CUDA total and insufficiently variable.