How to install CUDA in Linux?

Install the NVIDIA Driver and CUDA Toolkit

TIP

Traditionally, the NVIDIA Driver and CUDA Toolkit are installed separately, but you can actually install the CUDA Toolkit directly, and the system will automatically install the NVIDIA Driver that matches its version.

Following is the detailed installation steps, with CUDA-10.1 as an example:

1. Check if there is a GPU that supports CUDA

lspci | grep -i nvidia

2. Install the gcc and make

sudo apt update
sudo apt install gcc g++ make
# Install the dependent libraries required to run the CUDA example
sudo apt install libglul-mesa libxi-dev libxmu-dev libglul-mesa-dev freelut3-dev

Sometimes, too high versions of gcc and g++ may cause GPUMD compilation errors, so you need to download lower versions of gcc and g++ and switch versions (gcc is used as a demonstration below):

Check the current default gcc version: gcc --version
Check the installed gcc version: ls /usr/bin/gcc*
If there is no suitable version then required, perform a specified installation, such as: sudo apt-get install gcc-7
Then perform manual version management: sudo update-alternatives --config gcc, that’s it.

3. Download CUDA Toolkit

CUDA Toolkit

4. Run and install

The deb package can automatically install the NVIDIA driver by default (the preferred way).

Note: when installing Ubuntu, press E to pop up GRUB mode, add nouveau.modeset=0 at the end of the Linux line, and then press F10 to restart. Before installing Ubuntu dual system, please create a blank partition in advance.

WARNING: About Nouveau driver disabling

Generally, the deb package to install CUDA will automatically install the Nvidai driver, but there may still be problems in actual testing. This may be caused by the open source driver Nouveau being incompletely disabled.

Open the balcklist: sudo vim /etc/modprobe.d/blacklist.conf

Insert the following content at the bottom and save it:
blacklist nouveau
blacklist lbm-nouveau
options nouveau modeset=0
alias nouveau off
alias lbm-nouveau off
Update: sudo update-initramfs -u

Check after restart: lsmod | grep nouveau

5. Modify environment variables

sudo vim ~/.bashrc

Add the following environment variables:

export PATH=$PATH:/usr/local/cuda-10.1/bin
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-10.1/lib64
export CUDA_HOME=$CUDA_HOME:/usr/local/suda-10.1

6. Test

nvidia-smi
nvcc --version

nvcc is a CUDA C/C++ compiler that can directly compile source files (.cu) containing C++ syntax. The syntax is similar to gcc. Its path is located at /usr/local/cuda-10.1/bin

TIP