The following procedure will guide you through installation of tensorflow on AWS GPU instances with 64bit Ubuntu1604

make sure .bashrc has the following content

export CUDA_HOME=/usr/local/cuda
export CUDA_ROOT=/usr/local/cuda
export PATH=$PATH:$CUDA_ROOT/bin
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CUDA_ROOT/lib64

CUDA and cuDNN Installation

Install CUDA Toolkit 8.0

# note 1

wget <http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/cuda-repo-ubuntu1604_8.0.61-1_amd64.deb>
sudo dpkg -i cuda-repo-ubuntu1604_8.0.61-1_amd64.deb
sudo apt-get update

# note 2

sudo apt-get install -y nvidia-367 libcuda1-367
sudo apt-get install -y cuda-toolkit-8-0
sudo reboot

note 1

If your linux is not ubuntu1604, you should browse here to find your distribution and version of linux.

note 2

Currently AWS has two instance types with GPU support: G2 and P2. G2 is installed with K520 card and P2 is with P80. nvidia-367 driver is for the K520 card therefore the G2 instances. If you are using P2 instance, you should use nvidia-375 and libcuda1-375

Check driver state

nvidia-smi should show all your GPU cards now.

cuDNN v5.1

wget <https://dl.dropboxusercontent.com/u/2762665/cudnn-8.0-linux-x64-v5.1.tgz>
sudo tar -xvf cudnn-8.0-linux-x64-v5.1.tgz -C /usr/local

Tensorflow Installation

install

sudo apt-get install -y python3-pip
pip3 install tensorflow-gpu
pip3 install opencv-python

verify

import tensorflow as tf
tf_session = tf.Session()
x = tf.constant(1)
y = tf.constant(1)
tf_session.run(x + y)