0

I am not familar with CUDA so there might be some mistakse in my coments.

I am currently using RTX 3080 and CUDA 11.4.

For my project, CUDA 10.2 is recommended, So I used nvidia-docker and installed CUDA 10.2 but faced with this error.

ValueError: Unknown CUDA arch (8.6) or GPU not supported

I guess this is issue with RTX 3080 not suporting CUDA 10.2. So aren't there any ways to use CUDA 10.2 with RTX 3080?

Dongri
  • 1
  • 2
  • why CUDA10.2 ? you should use the latest version of CUDA, 11.4 I guess – Bruno Magacho Jul 23 '21 at 03:12
  • I want to use maskrcnn_benchmark but it requires CUDA 10.2 . (https://github.com/facebookresearch/maskrcnn-benchmark ) – Dongri Jul 23 '21 at 03:16
  • and it doesn't work with latest versions? – Bruno Magacho Jul 23 '21 at 03:17
  • I had some issues, but I will try it again and check the error again – Dongri Jul 23 '21 at 03:19
  • 2
    Off the shelf, you probably can't (see the linked duplicate for an explanation). You *might* be able to modify the build to an earlier architecture and let the driver perform JIT recompilation, but (a) Depending on the build system that might not be possible, (b) the JIT recompilation will have enormous performance impacts because it is very slow for complex codebases, and (c) it won't be able to use any of the most advanced/newest features of your GPU, so it wouldn't be much of a benchmarking exercise anyway – talonmies Jul 23 '21 at 03:20

0 Answers0