Nvidia CUDA
CUDA is a parallel computing platform and API model that allows developers to use CUDA-enabled GPUs for general purpose processing. CUDA works with all Nvidia GPUs from the G8x series onwards, including GeForce, Quadro and the Tesla line and is compatible with most standard operating systems.
Backwards and Fowards compatibility: https://docs.nvidia.com/pdf/CUDA_Compatibility.pdf, https://docs.nvidia.com/deploy/cuda-compatibility/
The CUDA toolkit consists of two main components:
The development libraries (including the CUDA runtime)
The driver components. The driver components are further separated into two categories:
Kernel mode components (the ‘display’ driver)
User mode components (the CUDA driver, the OpenGL driver, etc.)
Each release of the CUDA Toolkit requires a minimum version of the CUDA driver. The CUDA driver is backward compatible, meaning that applications compiled against a particular version of the CUDA will continue to work on subsequent (later) driver releases.
The NVIDIA driver can be installed as part of the CUDA Toolkit installation, but you should install it separately.
The CUDA driver on Linux systems is libcuda.so.
The minimum recommended GCC compiler (GNU Compiler Collection) is at least GCC 6.
Unless you are using a supercomputer you probably don’t have a POWER8 or POWER9 system, so ignore those steps. You likely have a x86_64 workstation. Only Tesla V100 and T4 GPUs are supported for CUDA 11.4 on Arm64 (aarch64) POWER9 (ppc64le).
It’s possible to install CUDA on WSL: https://docs.nvidia.com/cuda/wsl-user-guide/index.html
CUDA Installation
The steps in this section are mainly derived or copied straight from the official CUDA Linux guide and this article: https://www.iridescent.io/tech-blogs-installing-cuda-the-right-way/
1. Pre-installation Actions
Check your GPU compatibility
NVIDIA Clara needs a compute capability of 6.0 or higher. Your GPU should be listed here: https://developer.nvidia.com/cuda-gpus. You can find out what GPU you have with:
lspci | grep -I NVIDIA
Verify you have a supported version of Linux
The CUDA Toolkit is only supported on specific Linux distributions, which are listed in the release notes. Find your system information with:
uname -m && cat /etc/*release
x86_64 means you have a 64-bit system, and the remainder is your distribution information
Install GCC
Check if you have GCC by running gcc --version . It should be at least GCC 6.
If you don’t have gcc, install it using sudo apt install gcc
2. Install NVIDIA drivers
Method 1 (Using package manager as of Ubuntu 19.04)
Run nvidia-smi to verify your NVIDIA drivers are set up properly. If that doesn’t work, use modinfo nvidia to check your GPU.
If nvidia-smi didn’t work, install the latest version of nvidia-utils that is recommended in the terminal. For example: sudo apt install nvidia-utils-470
You can also check version using cat /proc/driver/nvidia/version
Method 2 (Only if method 1 doesn’t work)
Search for your GPU driver and download the Production Branch here: https://www.nvidia.com/Download/index.aspx?lang=en-us
Run chmod +x <driver file>, replacing <driver file> with the downloaded driver file name.
Reboot the computer and enter a TTY session by pressing Ctrl + Alt + F3 or another key combination if you’re not using a recent version of Ubuntu.
More about TTY here: https://askubuntu.com/questions/66195/what-is-a-tty-and-how-do-i-access-a-tty
Tip
Move between a few TTY terminals including the login screen so it’s not as buggy later on (this may be because you need to launch TTY from login screen)
Kill the X-server using sudo service lightdm stop, or if that doesn’t work, kill GDM using sudo service gdm stop.
If the screen goes blank, switch to another TTY terminal. If that still doesn’t work, reboot the computer, and try switching between TTY terminals beforehand.
Change directory to the location of the driver and run sudo ./NVIDIA-Linux-<system>-<version>.run
3. Download and install CUDA
Download the CUDA runfile (local) installer for your distribution from https://developer.nvidia.com/cuda-downloads and run sudo sh <downloaded file>, where <downloaded file> is the downloaded file. If the terminal says it is not executable, run chmod +x <downloaded file> first.
The runfile should be easier to use and more reliable than the deb files
If it says you already have a driver, continue anyways. Accept TOS. Select the Driver checkbox to uncheck it. Install.
Open your ~/.bashrc file (nano ~/.bashrc), and to the end of the file, append export PATH=$PATH:/usr/local/cuda/bin and export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64
Reboot and run nvcc --version to test installation success.
If the command is not found, you did something wrong, likely at the
bashrcstep. Do not install nvidia-cuda-toolkitYou shouldn’t need to reboot; you just need to run
source ~/.bashrc