CUDA profiling without GUI and then interpreting results with visual profiler

I want to be able to profile my CUDA code without being timed out by the X server. When I run some gpu code for more than 10 seconds, it times out. If I press CTRL+ALT+F1 I can disable the gui and run my cuda code there, but then I can’t run the NVIDIA compute visual profiler. I had a look at this post and it says I can profile without the GUI by setting the environment variable CUDA_PROFILE=1, but then the results are in a text file and can’t be read with visual profiler.

Is there any way to profile my CUDA code in a linux shell without a gui, and then interpret the results with the NVIDIA compute visual profiler?


If you have a second computer on the network with an X server, you should be able to ssh to the CUDA computer and run the visual profiler over X forwarding.