Cudatoolkit and NVCC


I have installed cudatoolkit=11.2 and cudnn=8.1 (from conda-forge channel) in my conda environment, and then I installed tensorflow=2.9. So I installed these specific versions of cudatoolkit and cudnn because tf2.9 requires those. So I am doing this in my conda env just to note. When I test tf, it recognizes my GPU, so there is no problem about that.

I am interested now, do I need to install NVCC? If so, when I try to install it from nvidia channel in conda, earliest version of cuda-nvcc is 11.3.58, but my cudatoolkit is version 11.2

Can someone explain me, if I have some cudatoolkit version, should NVCC version be > or <= of cudatoolkit version?

Hi @boigpu, By default when you install the cudatoolkit the nvcc will be automatically installed. It should be equal to the cudatoolkit version. Thank You.