How to install TensorFlow 1.13 with GPU support on Ubuntu 18.04 LTS + CUDA 10.0

8 min readDec 19, 2018

The purpose of this guide is to demonstrate how to install TensorFlow on Ubuntu 18.04 LTS using CUDA-enabled NVIDIA GPU’s. No prior experience with Ubuntu or TensorFlow is required.

Please note: In the following, all commands starting with a $sign (do not copy the dollar sign when copying commands) are expected to be executed within a command line interface tool such as Terminal (which can be opened on Ubuntu by pressing Ctrl+Alt+t).

Update (December 27, 2018): I added a troubleshooting section at the very bottom of the article. Please comment if you encounter other issues so I can add them to the troubleshooting section.

Update (April 26, 2019): This guide was written in December 2018. Since then, NVIDIA updated CUDA 10.0 to CUDA 10.1. However,TensorFlow does not yet support CUDA 10.1. Thus, please adjust all the download locations from NVIDIA to get the legacy 10.0 versions until Google updates TensorFlow.

Step 0: Prerequisites

Ubuntu 18.04 LTS. You can verify your Ubuntu version with the command:
$ lsb_release -a
CUDA-enabled GPU from NVIDIA. While nearly all newer NVIDIA GPU‘s are CUDA-enabled, please verify this for your specific GPU model here.
Installed GNU Compiler Collection (GCC). You can verify that you have GCC installed by issuing the command:
$ dpkg -s gcc
In case you get the message “ddpkg-query: package ‘gcc’ is not installed and no information is available …”, please install it with the command:
$ sudo apt update && sudo apt install gcc
Installed build-essential package. You can verify that you have the build-essential package installed by issuing the command:
$ dpkg -s build-essential
In case you get the message “dpkg-query: package ‘build-essential’ is not installed …”, please install it with the command:
$ sudo apt update && sudo apt install build-essential
Installed OpenGL Utility Toolkit (FreeGlut) and X11 Input extension library. To install these packages, please enter the following command:
$ sudo apt update && sudo apt install freeglut3 freeglut3-dev libxi-dev libxmu-dev
Installed pip3 and virtualenv. To install these packages, run the following commands subsequently:
$ sudo apt update && sudo apt install python3-dev python3-pip
$ sudo pip3 install -U virtualenv

Step 1: Installing NVIDIA’s Linux Display Driver via Runfile

Locate and Download NVIDIA’s Linux Display Driver: Navigate to NVIDIA’s Driver Download page, and enter the details of your specific GPU. Select Linux 64-bit as your operating system and click on the green search button (see Exhibit 1). On the following two pages, please click on the green download buttons and save the file to your ~/Downloads folder.

Exhibit 1: NVIDIA Driver Download Configuration (Example)

Install NVIDIA’s Display Driver: Once the file NVIDIA-Linux-x86_64–[YOURVERSION].run file is downloaded, please navigate to the ~/Downloads folder within the Terminal ($ cd ~/Downloads) and run the following shell script :

$ sudo sh ./NVIDIA-Linux-X86_64-[YOURVERSION].run

As noted by NVIDIA, at the end of the installation, you can either directly allow the modification of the X configuration file, do it manually or run $ nvidia-xconfig.My recommendation is to allow the X configuration modification by the installer.

Restart Machine and Verify: Once the Display Driver is installed, restart your system and verify the installation by executing the NVIDIA System Management Interface:

$ nvidia-smi

… which produces an overview of different metrics as shown in Exhibit 2.

Exhibit 2: Example Output of nvidia-smi Tool

Step 2: Installing CUDA Toolkit 10 via Runfile

“The NVIDIA CUDA Toolkit provides command-line and graphical tools for building, debugging and optimizing the performance of applications accelerated by NVIDIA GPUs, runtime and math libraries, and documentation including programming guides, user manuals, and API references” (Source: NVIDIA CUDA Toolkit 10 installer, 2018).

Locate and Download NVIDIA’s CUDA Toolkit 10: Head over to NVIDIA’s CUDA Toolkit Download page. The suitable selections should already be made (see Exhibit 3). Then, please click on the green download button and download the cuda_10.[YOURVERSION]_linux.run file to your ~/Downloads folder.

Install NVIDIA’s CUDA Toolkit 10 via Runfile: Once the file NVIDIA-Linux-x86_64–[YOURVERSION].run file is downloaded, navigate to the ~/Downloads folder and run the command:

$ sudo sh cuda_10.[YOURVERSION]_linux.run

To skip the license agreement, press Ctrl+c

IMPORTANT: During the installation, type “no” when asked if you want to “Install NVIDIA Accelerated Graphics Driver for Linux-x86_64 [YOURVERSION]?”. This is because, as of December 19th, 2018, the NVIDIA CUDA Toolkit 10 installer ships with an older NVIDIA Display Driver (version 410.48) than is offered on the NVIDIA Display Driver download page from step 1 (version 410.78). Therefore, the NVIDIA CUDA Toolkit 10 installer finishes with a warning message ***WARNING: Incomplete installation! This installation did not install the CUDA Driver. …”.

Setup the environment variables: To do so, please open the .bashrc file in the editor nano by entering:

$ sudo nano ~/.bashrc

Scroll down and add the following two paths to the .bashrc file:

export PATH=/usr/local/cuda-10.0/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda-10.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

To save the changes, please press Ctrl + o, press enterto accept the changes and Ctrl + xto close nano. Then, to reload the .bashrc file with the changes made, please enter the following command to the Terminal:

$ source ~/.bashrc

Verify the installation of NVIDIA’s CUDA Toolkit 10 compiler driver: To do so, please enter the following command (the output should look similar to Exhibit 4):

$ nvcc -V

Exhibit 4: Example of CUDA compiler driver install

Test the installation of NVIDIA’s CUDA Toolkit 10: For this, enter the following commands to compile the CUDA 10 samples, which will take a while:

$ cd ~/NVIDIA_CUDA-10.0_Samples
$ make

Once you see the message “Finished building CUDA samples”, enter the following commands to test CUDA GPU jobs with the deviceQuery program:

$ cd ~/NVIDIA_CUDA-10.0_Samples/bin/x86_64/linux/release
$./deviceQuery

The result should look similar to Exhibit 5. Note the Result = PASS at the very bottom. If you see this message, congratulations, you have successfully installed CUDA Toolkit 10!

Exhibit 5: Example Output of the deviceQuery Application

Step 3: Installing NVIDIA cuDNN

“The NVIDIA CUDA Deep Neural Network library (cuDNN) is a GPU-accelerated library of primitives for deep neural networks.” (Source: NVIDIA 2018)

Locate and Download NVIDIA’s cuDNN 7.4: Please head over to the NVIDIA cuDNN website and click on the green Download cuDNN button. Unfortunately, you have to register for the NVIDIA Developer Program. Once you are registered, go to the NVIDIA cuDNN website and fill in the survey to reach the download page (see Exhibit 6).

Here, please click on Download CuDNN v7.4-[CURRENTVERSION), for CUDA 10.0, and then download the following three .deb files for Ubuntu 18.04 to your ~/Downloads folder:

- cuDNN Runtime Library for Ubuntu18.04 (Deb)
- cuDNN Developer Library for Ubuntu18.04 (Deb)
- cuDNN Code Samples and User Guide for Ubuntu18.04 (Deb)

Install NVIDIA’s cuDNN 7.4: To install the three cuDNN packages, please execute the following commands one after another:

$ cd ~/Downloads
$ sudo dpkg -i libcudnn7_7.4.2.24-1+cuda10.0_amd64.deb
$ sudo dpkg -i libcudnn7-dev_7.4.2.24-1+cuda10.0_amd64.deb
$ sudo dpkg -i libcudnn7-doc_7.4.2.24-1+cuda10.0_amd64.deb

Verify the cuDNN installation: Please compile the MNIST example by executing the following commands as suggested by NVIDIA:

$ cp -r /usr/src/cudnn_samples_v7/ $HOME
$ cd $HOME/cudnn_samples_v7/mnistCUDNN
$ make clean && make
$ ./mnistCUDNN

If cuDNN is properly installed and running on your system, you will see a message “Test passed” as shown in Exhibit 7.

Exhibit 7: Verification of cuDNN installation with MNIST example

Step 4: Installing TensorFlow 1.13

TensorFlow is “an open source machine learning framework for everyone” (Source: Google 2018)

Create a Virtual Environment (optional but recommended): Virtual environments help you to isolate installed Python packages from the system. The following steps are adapted from TensorFlow installation guidelines which you can find here.

To create a new virtual environment for Python3 and store it in ./venv, execute the following:

$ virtualenv --system-site-packages -p python3 ./venv

Then, to activate the newly created environment, and also to activate it in the future, please execute:

$ source ./venv/bin/activate

Within the virtual environment, you can install packages without messing around with the host system itself. For example, you can upgrade pip and show the packages installed within the virtual environment by executing:

$ pip install --upgrade pip
$ pip list

Later, you can exit the virtual environment by entering $ deactivate (DO NOT EXIT THE VIRTUAL ENVIRONMENT NOW!)

Install TensorFlow: Finally, you can install the GPU-enabled version of TensorFlow with the command:

$ pip install tf-nightly-gpu

IMPORTANT: Note that we are installing the latest tf-nightly-gpu build, since the non-nightly GPU-version of TensorFlow can’t import “libcublas.so.9.0”. I will update this once this issue is fixed.

Verify TensorFlow installation: Please execute the following Python snipped to verify your TensorFlow installation:

$ python -c "from tensorflow.python.client import device_lib; print(device_lib.list_local_devices())"

Note: If you get the error “ModuleNotFoundError: No module named ‘pandas’”, please install it via:$ pip install pandas

In case of a successfully installed GPU-enabled TensorFlow version, the output of the above will entail your GPU (in my case the GeForce RTX 2070) as shown below:

physical_device_desc: "device: 0, name: GeForce RTX 2070, pci bus id: 0000:07:00.0, compute capability: 7.5"

Troubleshooting

Issues with NVIDIA Display Driver after updating Linux kernel (December 27, 2018). If you run into problems after upgrading your Linux kernel, please re-run step 1 of this tutorial to rebuild the kernel modules.
Errors when executing ./deviceQuery after successful building of CUDA samples (January 17th, 2019). Please reboot your system and try again.
Random lags / lagging mouse after installing NVIDIA display driver. Please sync your display device to one specific monitor by starting the nvidia-settings tool as root($ sudo nvidia-settings). Here, navigate to “X Server XVideo Settings”. Now select one of your physical monitors to sync the display device to (instead of Auto).
Random system reboots while running complex TensorFlow models. This could be due to spikes in energy consumption of your graphics card exceeding a maximum wattage threshold. Changing this threshold to a lower value can potentially mitigate this. See https://github.com/tensorflow/tensorflow/issues/8858.

Contributors

Danney Az, Z* (the superninja)

Epilog

I really hope this guide helps you to install TensorFlow with GPU-support on a freshly installed machine running Ubuntu 18.04 LTS with a CUDA 10 enabled NVIDIA GPU. Please let me know if you spot any bugs. I will do my best to incorporate them as fast as possible.

For comparison purposes, here are my hardware specifications: NVIDIA GeForce RTX 2070 GPU, AMD Ryzen 5 2600X CPU, 32 GB DDR4 RAM, Samsung 512 GB NVMe m.2 SSD.

Please support me on Patreon: https://www.patreon.com/cjanze