Do we really need two copies of cudaDNN files?

TensorFlow General Discussion
, ,
1

As per the instructions here, Overview — NVIDIA cuDNN v9.2.1 documentation,
all the cudaDNN files should be copied into the main cuda directory.

But the instructions at Install TensorFlow with pip say that both the main
cuda folder as well as the folder where cudaDNN was installed must be on the path.

Do we still need the folder to cudaDNN on path if all of the files in that directory
were copied to the main cuda directory? Can I delete one of the two sets of these repeated files?

3

Hi @CarsonW

Welcome to the TensorFlow Forum!

No, You dont need two copies of the CUDA and cuDNN. Single installation is enough for these libraries in your system to detect for the Tensorflow GPU support.