CUDA Error thrown on connection to process or launch

I am getting the error “CUDA Driver version is insufficient for CUDA runtime version” when using this profiler on my application." The error occurs right after launching the application, but before the CUDA API is initialized in my application.

I haven’t been able to find any logs about what could be causing this problem. I also cannot run the application through a non local-host connection on my machine either. In this case, the launch window pops up and immediately closes.

Currently using:

  • Nsight Compute 2019.1 CUDA 10.1. Display Driver 418.67 Centos 7
  • We are using NVIDIA RTX 2080 TIs in our computers.

    When running the code outside of the profiler it runs correctly with no errors. I have problems when launching it via nvvp which doesn’t appear to support the RTX cards at the moment. It worked previously when I had a GTX 1080.

    Thanks,

    Tim

    Found a problem. Apparently profiling on this driver now requires sudo access. I don’t remember this being the case in the past.
    ( When profiling with the 2080Ti instead of showing a warning for this, my code just crashes when profiling with a CUDA error. )

    After correcting the previous issue nvprof and nvvp works in a limited fashion with the RTX 2080. I still am getting the cudaDriverGetVersion error with this fix enabled when using Nsight Compute. Nsight Compute is also not working with the GTX 1060 as well.

    I have also found that this doesn’t work on the CUDA Samples either. Which run fine otherwise without the profiler.

    Thanks,

    Tim