I am trying to profile my cuda app using Nsight Compute. Having the newest toolkit installed this is Nsight Compute 2022.3.
But this gives me an error of a non supported gpu-chip, which is gp106 of my older GTX 1060 card.
Apparently Nsight Compute 2019.5.0 is the last one that supports that card.
I downloaded and installed that version, but it does never complete the profiling after launch. I assume the different toolkit versions do not match.
Do I have to uninstall now all new drivers, toolkit etc. and than reinstall an old version? I am not sure which software components relate.
I am using Visual Studio 2022, so I fear I cannot compile with an older nvcc compiler.
As long as I can profile from command line I am fine.
Nsight Compute supports Volta and newer GPUs. You can find this info here https://developer.nvidia.com/nsight-compute in the Supported NVIDIA GPU architectures section towards the bottom. I don’t believe it ever supported Pascal, which is GTX 1060 if I’m not mistaken. Are you seeing info somewhere that says Nsight Compute 2019.5.0 would support Pascal/GTX1060?
Thanks for pointing this out. We did have some initial support for Pascal, but the underlying performance monitoring technology dropped support shortly after that release. For Pascal, I would recommend trying the legacy tools nvprof and nvvp which still have Pascal support and are included in the latest CUDA Toolkits. In theory you could try the 2019 version with the newest CTK, but it’s not a tested environment. You could also try an old CTK, but there have been so many improvements since then, that I probably wouldn’t recommend it.