After intalling cuda-10.1, I realized that it doesn't work with tensorflow 1.13.1, so I had to downgrade to cuda-10.0.

I tried for a very long time, and at the end seem to have purged 10.1 (sudo apt purge cuda*), and then installed 10.0 successfully. However, even after reboot, nvidia-smi still comes up with version 10.1.

I removed cuda.list before the purge, as well as the cuda-10.1 directory under /usr/local.

nvcc --version now shows 10.0.

How is Ubuntu 16.04 hanging on cuda-10.1?

  • 1
    What is the output of nvcc -V? As far as I know the nvidia-smi reports that even if you don't have CUDA installed at all. – Terrance Jul 26 at 23:45
  • nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2018 NVIDIA Corporation Built on Sat_Aug_25_21:08:01_CDT_2018 Cuda compilation tools, release 10.0, V10.0.130. not sure I agree if you are saying to ignore what nvidia-smi shows. One evidence is that tensorflow 1.13.1 still can not detect the GPU. – bhomass Jul 27 at 9:27
  • I am running CUDA 10.1 and my nvidia-smi reads CUDA 10.2. If I uninstall and not have CUDA installed it still reads CUDA 10.2. This will probably help: stackoverflow.com/questions/53422407/… – Terrance Jul 27 at 14:06
  • ok, I see there is the possibility of nvidia-smi reporting a different cuda version. This does not move my installation forward however. I know the runtime version is in compliance with the tensorflow version, but tensorflow can't see the gpu still. – bhomass Aug 1 at 23:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.