What's the sampling interval when we profile with --system-profiling on ?


I am using nvprof with --print-gpu-tace and --system-profiling on for profiling.

I have 3 questions actually about this profiling command:

  1. if I don’t specify any function in my code for profiling, the profiling will start from the time my cuda driver API is initialized ? ( I don’t use the driver API in my code but as I understood, things happen implicitly)

  2. What’s the sampling interval, in the cuda profiler options (from profiler-users guide): continuous-sampling-intervaly default value is 2ms ! Isn’t this value ? because it’s not what I see in my profiling informations and it varies pretty much from one line to the other .

  3. What’s the accurancy of these system measuremnts, especially for power ?

Thank in advance for your help,


Ans #1:
Yes. It start from the first CUDA API call.

Ans #2:
The “–continuous-sampling-interval” is used in combination with “–event-collection-mode continuous” mode. Note that the --continuous-sampling-interval is not used for --system-profiling. The different types (power, clock, and thermal) are sampled at different rates; which are fixed and cannot be configured. The sampling points are not accurate


Thank you for your reply.

I launched my kernel for profiling, and what I saw from my power samples ( from wattmeter) that the nvprof is quite expensive in terms a power.

Since the launching of the command we can see +50W in the hole node power consumption.

Is this normal ??

Thank you for your reply