How much are nsys profiler overheads?


If i training my model with nsys profiler, find that training time is slightly increase.

Do you give me a information about profiler’s overhead? (per operation overhead or etc…)

Thanks you.

The overhead will vary depending on what options you are selecting and intensity of API usage. In general, we try to keep the overhead less than 5%.

If you could give me your command line (or which options you are selecting, if using the GUI) I could give you a more specific answer, and may be able to make recommendations for minimizing overhead.