ranr01/Dell_Precision_5520_install.md

## Dell_Precision_5520_install.md

      
    Raw
  

              Dell_Precision_5520_install.md
            
          
    Here we will set up a Ubuntu Dell Precision 5520 with Ubuntu 18.04, and
TensorFlow with GPU support.
Information collected from several sources cited below.


Upgrade to Ubuntu 18.04


Open a terminal and start with:
$ sudo apt update
$ sudo apt upgrade
$ sudo apt dist-upgrade
$ sudo apt-get autoremove
$ sudo apt install update-manager-core  


Make sure the last line in
/etc/update-manager/release-upgrades is set to Prompt=LTS


Finally do:
$ sudo do-release-upgrade


Install the NVIDIA driver.

Here I will setup the driver to a headless mode, i.e. the driver
will not handle the graphics of the computer (that will be left to
the integrated Intel GPU), but one can still use the NVIDIA GPU for
CUDA and TensorFlow.


Download the NVIDIA driver as a run file and make it executable
(chmod +x NVIDIA_installer.run)


Remove everything nvidia with apt-get. Something like
$ sudo apt-get purge nvidia

(maybe a good idea to restart after this one to make sure the
Intel graphics driver is working ok)


To install we need to run with no x-server so we change to
runlevel 3
$ sudo init 3


Log in to a console and install the driver
$ sudo ./NVIDIA_installer.run --no-drm --no-opengl-files

the ’no-drm’ makes the installer skip the automatic x-server
config and the ’-no-opengl-files’ prevents the replacement of
the system’s OpenGL files. If you want to use the driver for
graphics you should remove those.

The installer asks a bunch of questions and you should try and
answer them as best you can. I could not get the driver to build
using DKMS so I answered no to this question.


Restart and test that you have the driver with
$ nvidia-smi

which should show you the status of your discrete GPU.


Installing CUDA for Tensorflow


Download the CUDA toolkit from NVIDIA as a run file (if update
patches are available download them too). At the time of writing
the right CUDA version for TensoFlow 1.10 was CUDA 9.0 for
Ubuntu 17.04. It is important to use the CUDA version specified
in the TensorFlow documentation (unless you want to build TF
from source).


Run:
$ sudo chmod +x cuda_9.0.176_384.81_linux.run 
$ ./cuda_9.0.176_384.81_linux.run --override

say yes to installing with an unsupported configuration and
no to Install NVIDIA Accelerated Graphics Driver for
Linux-x86_64 384.81? (IMPORTANT!)


After the CUDA installation finishes add the following lines to
your ~/.bashrc file:
export PATH=/usr/local/cuda-9.0/bin:${PATH} 
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:${LD_LIBRARY_PATH)


run source ~/.bashrc to activate the installation


If there are published patches run-files, run them as well.


To test CUDA you can go into the code sample library and build
and run the deviceQuery utility.


Installing cuDNN


Download the right version of cuDNN from Nvidia (as specified in
the TensorFlow docs) and that is compatible with your CUDA
version.


Install with:
$ tar -zxvf cudnn-9.0-linux-x64-v7.tgz
$ sudo cp -P cuda/lib64/* /usr/local/cuda-9.0/lib64/ 
$ sudo cp  cuda/include/* /usr/local/cuda-9.0/include/
$ sudo chmod a+r /usr/local/cuda-9.0/include/cudnn.h


Finally install libcupti with
$ sudo apt-get install libcupti-dev


Install python and TensorFlow.

There are many ways to accomplish this. One of them is with
miniconda:
$ wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh 
$ bash Miniconda3-latest-Linux-x86_64.sh
# (accept the license and set a directory that works for you) 
$ conda create --name tf python=3.6  
# TF does not support python 3.7 at the time of writing
$ source activate tf 
(tf) $ pip install tensorflow-gpu


Sources
https://www.zdnet.com/article/how-to-upgrade-from-ubuntu-linux-16-04-to-18-04/
https://askubuntu.com/questions/1028949/why-am-i-not-getting-the-ubuntu-18-04-upgrade
https://medium.com/codezillas/step-by-step-guide-to-install-tensorflow-gpu-on-ubuntu-18-04-lts-6feceb0df5c0