Value 'sm_30' is not defined for option 'gpu-architecture'

[Qcellaris]

Hi,

I installed the cuda toolkit and worked through the tutorial where the function doublify was created to double all entries in an array. Unfortunately, I can't get it working. Running this code results in the following error message:

nvcc fatal : Value 'sm_30' is not defined for option 'gpu-architecture'

My setup:
Ubuntu 20.04
NVIDIA GeForce GTX 680
Driver Version: 470.82.01
CUDA Version: 11.4

Any clue if there is something wrong with my settings or is it something related to pycuda?

Kind regards,

[inducer]

I think the issue might be that Nvidia has removed support for the Kepler architecture from nvcc in the most recent versions. PyCUDA queries the GPU for its "compute capability", and it returned 3.0, which it then passes on to nvcc, which declared itself incapable of compiling for that compute capability. Downloading an older CUDA toolkit might help.

[Qcellaris]

Thank you for your support, I finally got it working. It took me several CUDA Toolkit installations before I found out that only CUDA Toolkits 9 and 10 support compute capability 3.0:

CUDA SDK 9.0 – 9.2 support compute capability 3.0 – 7.2
CUDA SDK 10.0 – 10.2 support compute capability 3.0 – 7.5

Hope this might help others that encounter this problem.