GPU Setup

Following the steps shown at Ubuntu 16.04 LTS, Tensorflow will use the CPU for all the computational problems. To make TensorFlow use your GPU, you need to do some more installation.

There is many ways to install TensorFlow / CUDA. This guide is only one of them.

As the writting of this documentation, TensorFlow GPU is officially supported for CUDA 9.0 with Nvidia drivers > 384.x and cuDNN >= 7.2

After these steps, you will have installed:

  • CUDA 9.0 and it’s dependencies
  • cuDNN 7.3.0 and it’s dependencies
  • TensorFlow with GPU support and it’s dependencies

Step 0 - Dependencies

You should have most of theses already.

$ sudo apt-get install build-essential cmake git unzip zip python-pip python3-pip python-virtualenv swig python-wheel libcurl3-dev curl python-dev python3-dev python-numpy python3-numpy
$ sudo apt-get install linux-headers-$(uname -r)

Step 1 - Cleanup

You need to make sure that you have nothing Nvidia or CUDA related installed on your machine.

You can follow theses steps if you want to uninstall CUDA, Nvidia and Tensorflow from your machine.

Do not worry, Nvidia drivers will be installed with CUDA later on.

  • Remove all Nvidia and CUDA related installation

Danger

Be careful, the following steps are destructive and will uninstall and remove any Nvidia drivers installed

$ sudo apt-get purge nvidia*
$ sudo apt-get purge cuda* # You may need to manually purge them, for example sudo apt-get purge cuda-cusparse-9-0
$ dpkg -l | grep '^rc' | awk '{print $2}' | grep cuda | sudo xargs dpkg --purge   # verify the output first so you don't delete something else...
$ dpkg -l | grep '^rc' | awk '{print $2}' | grep nvidia | sudo xargs dpkg --purge # verify the output first so you don't delete something else...
$ sudo apt-get autoremove
$ sudo apt-get autoclean
$ sudo rm -rf /usr/local/cuda*
  • Uninstall any TensorFlow installation
$ pip uninstall tensorflow
$ pip uninstall tensorflow-gpu
  • reboot!
$ sudo reboot

Step 1 - Install CUDA

You can download CUDA from Nvidia website and manually install it, but it is preferable to use their repository and install it using Ubuntu’s package manager.

  • Download and install CUDA 9.0
$ curl -O http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/cuda-repo-ubuntu1604_9.0.176-1_amd64.deb
$ sudo apt-key adv --fetch-keys http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/7fa2af80.pub
$ sudo dpkg -i ./cuda-repo-ubuntu1604_9.0.176-1_amd64.deb
$ sudo apt-get update
$ sudo apt-get install cuda-9-0 # this may take a while (~1.7G)
  • reboot!
$ sudo reboot
  • Verify installation
$ nvidia-smi # should return a list of GPUs with some metrics. Make sure the driver's version shown on the top is > 384.x
../../../_images/nvidia-smi_output.png 
$ nvcc -V # should return the CUDA compiler version installed. Make sure the version is 9.0

# example

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2017 NVIDIA Corporation
Built on Fri_Sep__1_21:08:03_CDT_2017
Cuda compilation tools, release 9.0, V9.0.176

If you do not pass any verification steps, go back to Step 1 - Cleanup.

Step 2 - Install cuDNN

Download cuDNN 7.3.0 for CUDA 9.0 from Nvidia’s cuDNN archive.

You may need to create a account if you do not have one yet.

  • Download and install
$ sudo tar -xzvf cudnn-9.0-linux-x64-v7.3.0.29.tgz
$ sudo cp cuda/include/cudnn.h /usr/local/cuda/include
$ sudo cp cuda/lib64/libcudnn* /usr/local/cuda/lib64
$ sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*
  • Update your bashrc.

In the case you have different CUDA version installed, change the folder to the CUDA version you want.

$ echo 'export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64"' >> ~/.bashrc
$ echo 'export CUDA_HOME=/usr/local/cuda' >> ~/.bashrc
$ echo 'export PATH="$PATH:/usr/local/cuda/bin"' >> ~/.bashrc
$ . ~/.bashrc

Step 3 - Install TensorFlow GPU

  • Uninstall TensorFlow
$ pip uninstall tensorflow
  • Install TensorFlow with GPU support under python3
$ python3 -m pip install --user tensorflow-gpu
  • Verify installation
$ python3
    import tensorflow as tf
    hello = tf.constant('Hello, TensorFlow!')
    sess = tf.Session() # You should see some information about your GPU in the output
    print(sess.run(hello))

# in another shell
$ nvidia-smi # you should see in the processe list python3

If you do not pass any verification steps, go back to Step 1 - Cleanup.

Step 5 - Profit

Have fun!