How to install CUDA and nvidia-driver

How to install CUDA and nvidia-driver

-** CUDA ** (Compute Unified Device Architecture) is a development environment for creating GPU-accelerated applications developed by Nvidia.

-** nvidia-driver **, as the name suggests, is a driver for controlling Nvidia GPUs such as the GTX and RTX series.

To briefly explain these two roles, ** nvidia-driver ** acts as an intermediary between hardware and OS, and a toolkit (application group) called ** CUDA ** is used for programs such as python. Provides resources.

Installation method

In the past, it was necessary to stop the running graphic driver, re-burn the kernel with dkms, etc., but today it can be easily installed with just one deb file.

Here are the steps to install CUDA 11.2 on Ubuntu 16.04.

1. Select your environment from the CUDA toolkit distribution site and check the .deb file download command.

First, access CUDA Toolkit 11.2 Downloads.

--If you select Linux-> x86_64-> Ubuntu-> 16.04-> deb (local)

Download Installer for Linux Ubuntu 16.04 x86_64

You will see how to download the .deb file in the content box.

2. Install CUDA using the deb file

The commands actually displayed are shown below, but since they are just examples, be sure to enter the commands displayed in the content box when you operated them yourself.

#Register priority in repository
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/cuda-ubuntu1604.pin
sudo mv cuda-ubuntu1604.pin /etc/apt/preferences.d/cuda-repository-pin-600

#Download deb file
wget https://developer.download.nvidia.com/compute/cuda/11.2.0/local_installers/cuda-repo-ubuntu1604-11-2-local_11.2.0-460.27.04-1_amd64.deb

#Install packages on your system using deb files
sudo dpkg -i cuda-repo-ubuntu1604-11-2-local_11.2.0-460.27.04-1_amd64.deb
sudo apt-key add /var/cuda-repo-ubuntu1604-11-2-local/7fa2af80.pub

#Updated package list
sudo apt-get update

3. Install the meta package that suits your environment

I was able to install the package with dpkg -i, but it has not been reflected in the system yet. Finally, use the sudo apt install command to apply the package to your system, but be careful here. In the content box, it says sudo apt install cuda, but you need to change the metapackage to reflect it according to your purpose.

The list of meta packages is excerpted from the official website (3.10.3. Meta Packages) and translated into Japanese below. 　 The following two meta packages are mainly specified.

• cuda-11-2 --When installing a new CUDA itself.
• cuda-toolkit-11-2 --When nvidia-driver is already installed --When you already have another CUDA and want to install another version of CUDA. --If you overwrite the driver poorly, the machine will stop working.

By the way, in the case of CUDA 10.1, the meta package name is cuda-10-1. Please convert it on a case-by-case basis.

sudo apt install cuda-11-2

Or

sudo apt install cuda-toolkit-11-2

Apply the CUDA package to your system with.

I want to manage multiple CUDA.

** CUDA ** and ** nvidia-driver ** are compatible. If you want to know more, please refer to the official website below.

Nvidia: Binary Compatibility

--If nvidia-driver == 410.48 is installed --CUDA 10.0, 9.2, 9.1 ... etc. can be installed, but CUDA11.0 cannot be installed. Please check in advance if the CUDA you are installing now is compatible with the driver.

By the way, I think about these things when I am in the following situation.

--When you can't or don't want to change the version of nvidia-driver --When working with Google colab --When installing CUDA additionally on a machine that has already built the environment

If you operate nvidia-driver poorly, you will end up destroying the environment and reinstalling the OS. Tedious.

CUDA installation location and how to switch

The reality of CUDA is in / usr/local. These are / usr/local/cuda11-2 and/usr/local/cuda10-2. However, programs such as python do not refer to them directly. Refer to a symbolic link called cuda.

Console

hoge@hoge:/usr/local$ ls -la
drwxr-xr-x 12 root root 4096 October 15 10:10 .
drwxr-xr-x 13 root root 4096 October 15 10:10 ..
～～～
lrwxrwxrwx 1 root root 9 October 15 10:10 cuda -> cuda-11.2
drwxr-xr-x 16 root root 4096 December 31 23:59 cuda-11-2
drwxr-xr-x 16 root root 4096 October 15 10:10 cuda-10-2
～～～

In the above situation, you can see that referencing the cuda directory is equivalent to referencing the cuda-11-2. If you want to change to cuda10-2, re-paste the link to cuda with the ln -nfs cuda-10-2 cuda command. By doing this, you can manage and operate multiple versions of CUDA.

installation location of cudnn

As an aside, a library called cudnn is located in /usr/include/cudnn.h. If you want to check these versions, you can check the versions by typing:

cat /usr/include/cudnn.h | grep CUDNN_MAJOR -A 2

cudnn depends on the version of CUDA. So if you change the version of CUDA, you need to uninstall the cudnn and then install the appropriate version.