GPU-Setup for Ubuntu 16.04/18.04

Provide some script for cuda-cudnn set up. Forked from this.

We provide scripts to generate cuda-cudnn script. Somehow, you need to download Code Samples and User Guide from nvidia's download site and put it in this folder, registration needed.

Usage:

• Generate Scripts with assigned version

python3 gen_sh.py --cuda 9.0 --cudnn 7.4 --ubuntu 1604

• Install CUDA

sudo chmod +x gpu-setup-part1-cuda.sh
./gpu-setup-part1-cuda.sh

It will reboot the machine once process completed. Have a drink, this may take 20~30 min, upto download time

• Install CUDNN

sudo chmod +x gpu-setup-part2-cudnn.sh
./gpu-setup-part2-cudnn.sh

This may take ~5 min

• Have a tensorflow test

python3 gpu-test.py

Footnotes:

• We use cuda 9.0 and cudnn 7.4 as basic example here. For the further combination, information can be find cudnn version, cuda version. Or there ubuntu1804 version.

• We currently only support the following cudnn-cuda pair:

CUDNN	CUDA
7.0	9.0
7.2	9.0
7.2	9.2
7.4	9.0
7.4	9.2
7.4	10.0
7.4	10.1
7.5	9.0
7.5	9.2
7.5	10.0
7.5	10.1

• Multiple CUDA environment These site1, site2 may helpful.

Once multiple cuda was installed, we may use the following code in bashrc/zshrc to control CUDA version.

export CUDA_HOME=/usr/local/cuda-9.0
export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$CUDA_HOME/extras/CUPTI/lib64
export PATH=$PATH:$CUDA_HOME/bin

Tensorflow 2.0

tensorflow 2.0 requires specific version of cuda/cudnn/nvidia-driver

Check this.