Yes, you can. The CUDA driver has built-in profiling tools. How to do this is discussed in the Compute_Profiler.txt file, which you will find in the doc directory of the toolbox, but the basic idea looks something like this:
$ COMPUTE_PROFILE=1 COMPUTE_PROFILE_CSV=1 COMPUTE_PROFILE_LOG=log.csv COMPUTE_PROFILE_CONFIG=config.txt ./app
which tells runtime to enable profiling, use csv format output written to log.csv , including profile statistics read from config.txt . After starting the application, the runtime will delete the output file with the results of the raw profiling. You can then use the tool as you wish to look at them. The visual profiler can be sure that it is open for output, but most of the fantastic synchronization that it does requires the result to be generated using its own profile configuration files (under the hood, it dynamically does the same thing as you manually). but to fly). I made several copies and copied copies of the configuration files so that I could regenerate specific application profile profiles without a profiler on the mute nodes. Not too much fun, but it can be done.