How to Install NVIDIA CUDA on Ubuntu 26.04

CUDA is NVIDIA’s platform for GPU-accelerated computing. On Ubuntu 26.04, Canonical added a cuda-toolkit meta-package to multiverse, so you can install a Ubuntu-managed 13.1.x toolchain without NVIDIA’s APT source first. When you need NVIDIA’s rolling 13.4 branch, add the ubuntu2604 network repository instead.

Either way you still need a proprietary NVIDIA driver that matches your GPU—the toolkit does not replace nvidia-smi.

Before you begin

Check Ubuntu version should show 26.04 before you enable multiverse or download ubuntu2604 packages. GPU-accelerated OpenCV builds that use -D WITH_CUDA=ON assume this stack—see install OpenCV on Ubuntu after nvcc works.

WSL2: This guide targets native Ubuntu. On WSL2, install the NVIDIA driver on Windows and follow CUDA on WSL.

Quick answer

Add export PATH=/usr/local/cuda/bin${PATH:+:${PATH}} to ~/.bashrc, then run nvcc --version.

Does your GPU support CUDA 13?

CUDA 13 dropped offline compilation for Maxwell, Pascal, and Volta. Turing and newer fit this page. Older GPUs need CUDA 12.x from the toolkit archive.

lspci | grep -i nvidia

Check or install the NVIDIA driver

cuda-toolkit does not install the driver on CUDA 13.4-style packaging.

nvidia-smi

When this works, do not reinstall the driver unless you need a newer branch.

If nvidia-smi is missing, refresh apt:

sudo apt update

See recommended drivers:

ubuntu-drivers devices

Install the recommended branch:

sudo ubuntu-drivers install

Install headers when DKMS builds a module:

sudo apt install -y linux-headers-$(uname -r)

Reboot after a first driver install:

sudo reboot

Method 1 — Ubuntu multiverse cuda-toolkit

Enable multiverse when apt cannot find the package:

sudo add-apt-repository multiverse

Refresh indexes:

sudo apt update

Install Ubuntu’s meta-package:

sudo apt install -y cuda-toolkit

Check the candidate:

apt-cache policy cuda-toolkit

Candidate: 13.1.1-0ubuntu1
 13.1.1-0ubuntu1 500
        500 http://archive.ubuntu.com/ubuntu resolute/multiverse amd64 Packages

If NVIDIA’s repository is also enabled, APT may prefer 13.4.x. Pin Ubuntu’s build:

sudo apt install -y cuda-toolkit=13.1.1-0ubuntu1

Method 2 — NVIDIA ubuntu2604 repository

Use this for CUDA 13.4.x or version pins. Map architecture:

dpkg --print-architecture

Download the keyring with wget:

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2604/x86_64/cuda-keyring_1.1-1_all.deb

Install it:

sudo dpkg -i cuda-keyring_1.1-1_all.deb

Update apt:

sudo apt update

Install the toolkit meta-package from NVIDIA’s repo:

sudo apt install -y cuda-toolkit

Confirm apt selects the NVIDIA build:

apt-cache policy cuda-toolkit

Candidate: 13.4.3-1
 13.4.3-1 600
        600 https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2604/x86_64  Packages

Pin 13.4 when frameworks require that minor:

sudo apt install -y cuda-toolkit-13-4

Set the CUDA PATH

Append toolkit binaries to your user PATH:

echo 'export PATH=/usr/local/cuda/bin${PATH:+:${PATH}}' >> ~/.bashrc

Reload:

source ~/.bashrc

The /usr/local/cuda symlink should point at the version you installed:

ls -l /usr/local/cuda

Verify the toolkit

Ask the compiler for its release string:

nvcc --version

nvcc: NVIDIA (R) Cuda compiler driver
Cuda compilation tools, release 13.4, V13.4.xxx

On a GPU host, the driver line in nvidia-smi should meet the toolkit’s minimum:

nvidia-smi

Compile a CUDA test program

Install build tools:

sudo apt install -y build-essential

Create vector_add.cu with a simple __global__ kernel and cudaMalloc/cudaMemcpy (see NVIDIA CUDA samples for full examples). Compile on a GPU host:

nvcc -o vector_add vector_add.cu

Run:

./vector_add

Sample z[0] = 3.000000 (expect 3.0)

Update, pin, or uninstall

Refresh indexes:

sudo apt update

Pin or upgrade the 13.4 line:

sudo apt install cuda-toolkit-13-4

Remove toolkit packages:

sudo apt purge -y cuda-toolkit

Drop NVIDIA’s CUDA apt source when you no longer need updates from it:

sudo apt purge -y cuda-keyring

Run apt-cache policy cuda-toolkit when both Ubuntu and NVIDIA sources are enabled so you know which origin wins.

Troubleshooting

Summary

Ubuntu 26.04 can install cuda-toolkit from multiverse for a Ubuntu-managed 13.1 line, or add NVIDIA’s ubuntu2604 repository for 13.4.x. Install or verify the NVIDIA driver with nvidia-smi separately, set PATH, and confirm with nvcc --version.

Match GPU generation to toolkit major (CUDA 13 for Turing and newer). For 24.04 or 22.04, use the dedicated LTS guides in the intro table.

References

• CUDA Installation Guide for Linux
• CUDA downloads (Ubuntu 26.04 selector)
• Ubuntu 26.04 release notes — CUDA in archive
• CUDA toolkit archive