New machine instructions for GPU

Install the drivers

  • Download NVIDIA Drivers from here
    • Do not use apt-get install nvidia-* for drivers.
    • Generally they are few cycles older than those on the website.
  • You will get a file named something like this NVIDIA-Linux-x86_64-375.39.run
  • Give it executable permission
    • chmod +x NVIDIA-Linux-x86_64-375.39.run
  • Run with sudo
    • sudo ./NVIDIA-Linux-x86_64-375.39.run

Possible issues:

You might have to stop the X-server (the GUI) to install the drivers.

Use the following command on Ubuntu

sudo service lightdm stop

Install CUDA

  • Download the latest CUDA drivers from here

    Download the ‘runfile (local)’. This avoids issues of proxy while installing using network based debian file.

  • Use the following command to run the installer:

    • sudo sh cuda_8.0.61_375.26_linux.run
    • It will also prompt to install OpenCL and NVIDIA-Samples. It is not required but recommended to install those.
    • If you install NVIDIA-Samples, use 1_Utilities/deviceQuery to check the GPU specs.
    • It will ask you if you want to make soft-link to /usr/local/cuda - say YES

Install CuDNN

  • Download the latest CuDNN from here
  • Extract the tar file.
  • Copy all the files inside lib64/ to /usr/local/cuda/lib64/
  • Copy the file include/cudnn.h to /usr/local/cuda/include/
  • Some libraries will require LD_LIBRARY_PATH and LIBRARY_PATH to point to this lib64 directory:
    • export LD_LIBRARY_PATH=<path_to_lib64>:$LD_LIBRARY_PATH
    • export LIBRARY_PATH=<path_to_lib64>:$LIBRARY_PATH

Install Tensorflow

For stable release version:

  • sudo pip install –upgrade tensorflow_gpu

For latest nightly release:

  • sudo pip install –upgrade Get this from https://github.com/tensorflow/tensorflow