0

I was having issues with tensorflow not being able to detect any GPU (currently running version 2.4.0, and I would be willing to downgrade but need at least 2.0). Upon investigation of my cuda version I am getting conflicting results. Also, I looked up the driver version that I would have with my GPU (Titan X Pascal) here and it says it only supports version 430.64, but the nvidia-smi output is higher.

$nvidia-smi
    Mon Feb 22 10:05:09 2021       
    +-----------------------------------------------------------------------------+
    | NVIDIA-SMI 450.102.04   Driver Version: 450.102.04   CUDA Version: 11.0     |
    |-------------------------------+----------------------+----------------------+
    | GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
    | Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
    |                               |                      |               MIG M. |
    |===============================+======================+======================|
    |   0  TITAN X (Pascal)    On   | 00000000:83:00.0 Off |                  N/A |
    | 23%   24C    P8     9W / 250W |    139MiB / 12196MiB |      0%      Default |
    |                               |                      |                  N/A |
    +-------------------------------+----------------------+----------------------+
    |   1  TITAN X (Pascal)    On   | 00000000:84:00.0 Off |                  N/A |
    | 23%   22C    P8     9W / 250W |    139MiB / 12196MiB |      0%      Default |
    |                               |                      |                  N/A |
    +-------------------------------+----------------------+----------------------+

But here is the nvcc output.

$nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2017 NVIDIA Corporation
Built on Fri_Nov__3_21:07:56_CDT_2017
Cuda compilation tools, release 9.1, V9.1.85

I'm quite confused and any help would be appreciated.

Himani
  • 69
  • 9
  • TF 2.4 normally [expects/requires CUDA 11.0](https://www.tensorflow.org/install/source#gpu). No other CUDA version can be substituted. The linked duplicate explains the difference between your `nvcc` and `nvidia-smi` output, and explains what you can do to harmonize things. Ultimately you want to install CUDA 11.0 (you may have done that already) and *then set your PATH variable properly* (which is not correctly done, yet). – Robert Crovella Feb 22 '21 at 15:30
  • Thank you @RobertCrovella. Is setting the PATH variable in the linked post? The issue is I am not sure that my graphics card can support CUDA 11.0 because the website says it only has up to driver version 430.64 but the minimum for CUDA 11.0 requires 450.x or higher. But running nvidia-smi shows Driver Version: 450.102.04. So the conflicting information is confusing. – Himani Feb 22 '21 at 17:19
  • Your graphics card can support any recent version of CUDA up through the latest. Yes, the linked article discusses setting the PATH variable. – Robert Crovella Feb 22 '21 at 17:24

0 Answers0