How to manage different versions of CUDA, CuDNN and Nvidia Drivers?

Question

I have recently update GPU in my PC, running Ubuntu 14.04

I'm trying to run some benchmarks with Caffe, but they will strongly depend on version of CUDA, CuDNN and Nvidia Drivers.

Is there any option to manage different versions CUDA, CuDNN and Nvidia Drivers?

Will different versions interfere with each other? Is it possible to switch between versions?

Installed Nvidia drivers:

apt-cache search nvidia | grep version

nvidia-173 - NVIDIA legacy binary driver - version 173.14.39
nvidia-304 - NVIDIA legacy binary driver - version 304.131
nvidia-304-updates - NVIDIA legacy binary driver - version 304.131
nvidia-340 - NVIDIA binary driver - version 340.96
nvidia-340-updates - NVIDIA binary driver - version 340.96
nvidia-352-updates - NVIDIA binary driver - version 352.63
nvidia-352 - NVIDIA binary driver - version 352.79
nvidia-355 - NVIDIA binary driver - version 355.11
nvidia-358 - NVIDIA binary driver - version 358.16
nvidia-361 - NVIDIA binary driver - version 361.45.18
nvidia-364 - NVIDIA binary driver - version 364.19
nvidia-367 - NVIDIA binary driver - version 367.44
nvidia-370 - NVIDIA binary driver - version 370.23

Solution

I can only answer a part of your question.

I wrote a script to switch between multiple cuDNN versions. It currently supports v5.1, v6.0 and v7.0. You can find the script here.

Once you complete the installation and download the cuDNN versions you want to use, you can switch between v5.1, v6.0 and v7.0 with the click of a button.

This is what the script does:

#!/bin/bash

rm -f /usr/include/cudnn.h
rm -f /usr/lib/x86_64-linux-gnu/*libcudnn*
rm -f /usr/local/cuda-*/lib64/*libcudnn*


cp -P packages/cudnn/include/cudnn.h /usr/include
cp -P packages/cudnn/lib64/libcudnn* /usr/lib/x86_64-linux-gnu/
chmod a+r /usr/lib/x86_64-linux-gnu/libcudnn*

rm -rf packages/cudnn

It performs the installation operations automatically for the version you need.

Comments

[Jerry Yang]

About coexistence and management of different CUDA versions, I think you can find your answer here: devtalk.nvidia.com/default/topic/493290/…