Hi,
I find the tegrastats very problematic to say the least. You can see 100% in the GPU stats when your code is actually written badly - performance wise.
I found the easiest way to analyze the performance, is to run my application on Xavier like this:
nvprof -f -o xx.nvvp
Open the resulting .nvvp file on a desktop/laptop (i.e. not on Xavier) with nvvp and see the timeline, kernel statistics, kernel names etc…
Eyal