CUDA 11.0 compatible with Tensorflow 2.2?

Is CUDA 11.0 compatible with Tensorflow 2.2? And if not, why is it the only CUDA version which can easily be downloaded from the official Nvidia website?

I am asking, because I keep getting the error, that some dynamic libraries do not exist, when trying to run Tensorflow with GPU support in PyCharm. The dynamic libraries are:

  • libcudart.so.10.1
  • libcufft.so.10
  • libcurand.so.10
  • libcusolver.so.10
  • libcusparse.so.10

This issue is described here, where it is suggested to install CUDA 10.1 for Tensorflow 2.2. However I only find instructions on how to downgrade to CUDA 10.1 on unofficial websites like here. Why is there no official instructions from Nvidia where I can be sure that I do not break anything?

I followed all the Post-installation Actions. Also I set the environment variables ‘LD_LIBRARY_PATH=/usr/local/cuda-11.0/lib64’ in Pycharm. So the problem has to lie somewhere else.

1 Like