Why Installing cuDNN From Package Manager Doesn't Required Nvidia Developer Account

I am trying to install cuDNN library for tensorflow.

On the cuDNN installation guide from Nvidia (https://docs.nvidia.com/deeplearning/sdk/cudnn-install/index.html) it looked like I need to register for an Nvidia developer program account then download the file manually.

However onthe tensorflow installation guide (https://www.tensorflow.org/install/gpu), I only need to call the package from the package manager (apt install libcudnn7) without signing up for the Nvidia developer.

Why is there such a difference? Am I missing something?


In general, for the most up to date CUDA/CUDNN/etc. versions as well as for installing archived versions of CUDA/CUDNN/etc, the NVIDIA developer page is the place to go. It usually offers multiple installation options such as .deb, .run, tar.gz, network repo, etc. for different workflows.

I don’t think libcudnn is in the Ubuntu package repos by default. I think you have to add the network repo first, similar to what’s mentioned here: https://docs.nvidia.com/deeplearning/sdk/cudnn-install/index.html#network-installation

The step to add the network repos are actually listed on that TF page as well here with the wget commands, which they are pulling from the NVIDIA Developer page: https://www.tensorflow.org/install/gpu#ubuntu_1804_cuda_10

There may have been some exceptions made for login to allow for easier TF installation, but in general you can just go through the Developer program to get access to all of the options.