How to Install NVIDIA CUDA on Ubuntu 24.04

CUDA is NVIDIA’s platform for GPU computing. On Ubuntu 24.04 you need a working proprietary NVIDIA driver plus the CUDA Toolkit (nvcc, headers, and libraries). NVIDIA qualifies 24.04 in the CUDA Linux installation guide under the repository codename ubuntu2404.

This page walks through NVIDIA’s recommended network repository install (cuda-keyring + cuda-toolkit). Ubuntu’s older multiverse package nvidia-cuda-toolkit remains if you deliberately want a distro-pinned 12.0-era toolchain instead of current CUDA 13.

Before you begin

Confirm the release with check Ubuntu version — you should see 24.04 before you download ubuntu2404 packages. You need a CUDA-capable NVIDIA GPU on bare metal or supported passthrough; CPU-only hosts can install toolkit packages for cross-builds, but nvidia-smi and device kernels need hardware.

Install kernel headers before driver DKMS builds:

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

WSL2: Use NVIDIA’s CUDA on WSL guide—install the Windows driver on the host, not inside the Linux guest alone.

Quick answer

Add NVIDIA’s CUDA repository, install cuda-toolkit, and put /usr/local/cuda/bin on your PATH:

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb

Register the keyring:

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

Refresh indexes:

sudo apt update

Install the toolkit meta-package:

sudo apt install -y cuda-toolkit

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

Does your GPU support CUDA 13?

CUDA 13 dropped offline compilation support for Maxwell, Pascal, and Volta. Turing and newer (RTX 20xx and later, recent data-center GPUs) fit the CUDA 13 workflow on this page. Older silicon should use CUDA 12.x from NVIDIA’s toolkit archive.

List visible NVIDIA devices:

lspci | grep -i nvidia

No output usually means the OS does not see an NVIDIA PCI device.

Check or install the NVIDIA driver

From CUDA 13.4 packaging, cuda-toolkit does not install the NVIDIA driver. Verify the driver first:

nvidia-smi

When this prints GPU name, driver version, and a process table, skip driver reinstall unless you need a newer branch for your toolkit.

If nvidia-smi is missing, refresh apt and list recommendations:

sudo apt update

Ubuntu’s driver helper lists recommended branches for your hardware:

ubuntu-drivers devices

Install Ubuntu’s recommended branch (example output may show nvidia-driver-580):

sudo ubuntu-drivers install

Reboot after a first driver install:

sudo reboot

Install CUDA from NVIDIA’s ubuntu2404 repository

NVIDIA documents network repository setup in the CUDA installation guide — Ubuntu. Map architecture before you download the keyring:

dpkg --print-architecture

On amd64, fetch the keyring with the wget command:

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb

Install it with dpkg:

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

Update apt:

sudo apt update

Install the rolling toolkit meta-package:

sudo apt install -y cuda-toolkit

See which version apt will take:

apt-cache policy cuda-toolkit

Candidate: 13.4.3-1
 13.4.3-1 600
        600 https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64  Packages

Pin CUDA 13.4 when frameworks expect that minor:

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

NVIDIA installs under /usr/local/cuda-<version>/ with a symlink at /usr/local/cuda.

Optional Ubuntu multiverse nvidia-cuda-toolkit

Ubuntu 24.04 still publishes nvidia-cuda-toolkit in multiverse (roughly 12.0.x—check apt-cache policy nvidia-cuda-toolkit on your host). Use it only when you need Ubuntu’s pinned distro build instead of NVIDIA’s CUDA 13 line. Do not mix conflicting meta-packages without a deliberate purge plan.

Set the CUDA PATH

Append toolkit binaries to your user PATH:

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

Reload the shell:

source ~/.bashrc

Confirm the default symlink:

ls -l /usr/local/cuda

Verify the CUDA toolkit

Print the compiler version:

nvcc --version

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

Exact V13.4.* build strings change with package updates.

Compare driver support:

nvidia-smi

The CUDA version line in nvidia-smi should meet the minimum for your toolkit major—see CUDA release notes.

Compile a small CUDA test (GPU required)

Install a C++ toolchain:

sudo apt install -y build-essential

Save a minimal vector_add.cu (same pattern as NVIDIA’s getting-started samples), compile with nvcc -o vector_add vector_add.cu, and run ./vector_add on a host where nvidia-smi already works. Expect output similar to Sample z[0] = 3.000000 when the driver and toolkit align.

Update, pin, or remove CUDA

Refresh and upgrade the meta-package you use:

sudo apt update

Move to the 13.4 meta-package when you need that minor:

sudo apt install cuda-toolkit-13-4

List installed CUDA packages:

dpkg -l | grep -E '^ii\s+cuda'

Remove the toolkit meta-package when uninstalling:

sudo apt purge -y cuda-toolkit

Purge cuda-keyring if you no longer want NVIDIA’s repository:

sudo apt purge -y cuda-keyring

Purging cuda-toolkit does not remove nvidia-driver-* on CUDA 13.4-style packaging.

Troubleshooting

Summary

Ubuntu 24.04 installs current CUDA 13 through NVIDIA’s ubuntu2404 network repository: cuda-keyring, sudo apt install cuda-toolkit, then set PATH and verify with nvcc --version. Install or confirm the NVIDIA driver separately—cuda-toolkit does not pull a driver on CUDA 13.4-style packages.

Use cuda-toolkit-13-4 when you need a fixed 13.4 minor; use Ubuntu’s nvidia-cuda-toolkit only when you intentionally want the older multiverse toolchain. For 26.04 archive installs or 22.04 repo paths, see the sibling guides linked in the intro table.

References

• CUDA Installation Guide for Linux
• CUDA downloads (Ubuntu 24.04 selector)
• CUDA-supported GPUs
• CUDA toolkit archive