2

In the below image tensorflow-2.7 has CUDA version as 11.2 does that mean I can installed

  • versions > 11.2 or
  • versions < 11.2 or
  • must installed 11.2 only.

enter image description here

Vishal Reddy
  • 106
  • 3
  • 11

1 Answers1

4

Is the CUDA version mentioned in the list, the maximum version that can be installed or minimum version that should be installed?

Neither. It is the exact minor version of the CUDA runtime you must have available. Same for CUDNN (and that CUDNN version must also be version matched to your CUDA runtime).

In your example:

  1. versions > 11.2 or
  2. versions < 11.2 or
  3. must install 11.2 only.

it is the third option and only the third option which is correct. Note that if you see versions of CUDA such as 11.2.1 and 11.2.2, the last digit just indicates a bug fix revision, and there is no difference as far as Tensorflow is concerned.

talonmies
  • 70,661
  • 34
  • 192
  • 269
  • can I install 11.2.2 or should it be 11.2.0 exactly? what happens if i have 11.5 version – Vishal Reddy Jan 13 '22 at 08:06
  • 1
    Minor (third digit) versions are just bug fixes and that doesn't matter. If you have an 11.5 runtime for a Tensorflow build expecting 11.2, it will fail to run with missing library errors – talonmies Jan 13 '22 at 08:34
  • As for the version compat. between Tensorflow and cuDNN/CUDA: there don't appear to be hard and fast rules. You should see the table as the published results of successful builds. All they are saying is that, we compiled this version of TF with this version of cuCNN / CUDA and it worked. It is an 'empirical' result that was obtained from a particular combination of Pyhton, gcc, Bazel, cuDNN and CUDA. – darbehdar Mar 06 '23 at 07:12
  • 1
    As for cuDNN and CUDA version compatibility check out the cuDNN archive at https://developer.nvidia.com/rdp/cudnn-archive. You get a list of cuDNN versions with their corresponding CUDA version. You can see for example that there is a cuDNN v8.7 for CUDA 11.x as well as one for CUDA 10.2. If your CUDA is 10.2 make sure to open that link and download that version of cuDNN 8.7.0 – darbehdar Mar 06 '23 at 07:17