Linux with Tensorflow in MSI GL65 (with RTX2070)

I have been a linux user for a few years now.

I have also been interested in playing games in my computer (although being a gamer nowadays feels like a full-time job, so I don’t consider myself a hardcore gamer).

I often run a dual boot system, doing most of my daily chores on Linux, and playing during the weekends on my MS Windows partition.

To my surprise, installing a linux partition in this particular laptop was straightforward. No special settings on the bootloader, no special proprietary drivers for my WiFi. It just worked out of the box after I booted and installed a live usb flash drive.

Of course, the proprietary graphics driver, CUDA, and cuDNN have to be installed independently after the fresh OS install if you are interested in using the graphics card to train some deep learning models.

Those components are what I will be covering in this post.

I will also be assuming that you are running an Ubuntu 20.04 distro.

The steps should be fairly similar for other Debian-based distros.

Installing proprietary NVIDIA drivers

You have a few options for installing the graphic drivers. If you don’t feel like modding your default OS too much, I recommend you to simply install them from the Additional Drivers tab in your Software & Updates menu in Ubuntu.

Launch the Activities menu and type Additional Drivers

You should see the option for your NVIDIA RTX2070 there.

Install the latest version available and tested.

In my case, and at the time of this writing, that is nvidia-driver-450.

Installing CUDA

Tensorflow works with CUDA v10.1.

The CUDA toolkit for v10.1 can be found in the package manager.

sudo apt-get install nvidia-cuda-toolkit

Check the version

$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Sun_Jul_28_19:07:16_PDT_2019
Cuda compilation tools, release 10.1, V10.1.243

If you are in v10.1, you are good to go.

Installing cuDNN

Tensorflow requires cuDNN v7.6.

You can obtain the library either as a .deb package or by downloading the zipped .so libraries.

At the time of this writing, the .deb package is only available for Ubuntu <=18.04 and the folder structure differs from the folder structure of my cuda library installed from the package manager.

I suggest downloading the .so libraries.

You can either download the zipped files of v7.6.5, or browse the archive.

Note that you need to make an account to download these zipped files from NVIDIA

Uncompress the file somewhere in your file system

$ tar xvf cudnn-10.1-linux-x64-v7.6.5.32.tgz 

That should generate a folder named cuda wherever you uncompressed it.

For your convenience, make a variable CUDNNDIR that points to wherever place you uncompressed the file in.

In my case, I uncompressed inside my Download folder, so this is how it looks

$ CUDNNDIR=~/Downloads/cuda

This is the structure of my uncompressed file

$ tree $CUDNNDIR

.
├── include
│   └── cudnn.h
├── lib64
│   ├── libcudnn.so -> libcudnn.so.7
│   ├── libcudnn.so.7 -> libcudnn.so.7.6.5
│   ├── libcudnn.so.7.6.5
│   └── libcudnn_static.a
└── NVIDIA_SLA_cuDNN_Support.txt

In my Ubuntu 20.04 and using the nvidia-cuda-toolkit package from aptitude, the library is installed here:

CUDADIR=/lib/cuda

I also suggest making a variable that points to that place. Replace /lib/cuda with your location if that’s not where the library was installed.

The structure of my CUDADIR:

$ tree $CUDADIR

/lib/cuda
├── bin
├── include
├── lib64
├── nvvm
│   └── libdevice -> ../../nvidia-cuda-toolkit/libdevice
└── version.txt

The CUDADIR is where we need to install the cuDNN library.

Installing, in this context, is basically copying the headers and shared library. Then, making these files available to external libraries like tensorflow.

Copy the headers

sudo cp ${CUDNNDIR}/include/*.h ${CUDADIR}/include/

Copy the shared libraries

sudo cp ${CUDNNDIR}/lib64/libcudnn* ${CUDADIR}/lib64

The file structure of the CUDA library should look something like this now

/lib/cuda
├── bin
├── include
│   └── cudnn.h
├── lib64
│   ├── libcudnn.so
│   ├── libcudnn.so.7
│   ├── libcudnn.so.7.6.5
│   └── libcudnn_static.a
├── nvvm
│   └── libdevice -> ../../nvidia-cuda-toolkit/libdevice
└── version.txt

Change permissions to the recently copied files

sudo chmod a+r ${CUDADIR}/include/cudnn*.h ${CUDADIR}/lib64/libcudnn*

That installed cuDNN in your system.

At least in my case, these files are still not reachable by tensorflow. You can achieve that by exporting the path to your LD_LIBRARY_PATH variable.

export LD_LIBRARY_PATH=${CUDADIR}/lib64:$LD_LIBRARY_PATH

Now check whether tensorflow 2 is able to see your gpu

$ python3
Python 3.8.5 (default, Jul 28 2020, 12:59:40) 
[GCC 9.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
>>> tf.config.list_physical_devices('GPU') 
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

If you get a non-empty list as the result of list_physical_devices('GPU'), that should be a good indicator.

The export instruction is not persistent and the changes will be lost after you close your terminal.

Make it permanent by adding it to your ~/.bashrc file.

Putting new statements at the end of the file is usually the best practice.

When you do that, spell the $CUDADIR folder as an absolute path, do not use the $CUDADIR variable.

You should be all set. Try training something and check the console to corroborate it is doing it through your RTX2070.

Cheers.