Which Pytorch/Cuda version to use with newest 3060ti driver?

1

Hello,

I am running WSL-2 with Ubuntu-20.0.4
I’ve just installed a new 3060ti graphics card, and installed the latest driver with CUDA 11.6.
I was running CUDA 10.2 on my old GTX 1060 with no issues before.
See the nvidia-smi output below

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 510.47.03    Driver Version: 512.15       CUDA Version: 11.6     |
|-------------------------------+----------------------+----------------------+

I installed pytorch + cuda toolkit 11.3 in a conda environment as suggested on the Pytorch website

conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

However when running a python program, executing torch.cuda.is_available() returns False:

CUDA is available: False
CUDA toolkit version:  11.3
No GPU available, using the CPU instead.

Can anyone help walk me through why I am having issues? I figured the newest 512.15 graphics driver would be backwards compatible with CUDA 11.3, but I may be wrong.

2

I’m having exactly the same problem. Was this issue solved?

3

I’ve a nvidia 3060ti with Driver Version: 512.78.01 and CUDA 12.0.
Has anyone solved the problem of torch GPU detection?