# Install NVIDIA drivers and CUDA on Ubuntu or Debian

> Install the NVIDIA data center driver on a GPU server with ubuntu-drivers or NVIDIA's repository, add the CUDA toolkit, and confirm with nvidia-smi and nvcc.

Difficulty: Intermediate

Your GPU server's NVIDIA GPU needs the NVIDIA **driver** before any software can use it. Most AI tools and containers need nothing else; the **CUDA toolkit** is for compiling CUDA code yourself. This guide installs both the way Canonical, Debian and NVIDIA document it.

## Before you start

- A GPU VPS, GPU VDS or GPU dedicated server with a **dedicated NVIDIA GPU**. For a shared vGPU, see the note in [GPU servers: getting started](/guides/gpu-servers-getting-started).
- Ubuntu 24.04 LTS or 26.04 LTS, or Debian 12 or 13, fully updated and rebooted:

```bash
sudo apt update && sudo apt full-upgrade
sudo reboot
```

- Check that the GPU is visible: `lspci | grep -i nvidia`.
- Check Secure Boot: `mokutil --sb-state` (install `mokutil` if needed).

## Step 1: Install the driver

**Ubuntu**

Canonical's tool picks the right data center driver and installs pre-built, signed modules. List the candidates, then install one branch, for example the 535 branch Canonical uses in its documentation; pick a branch from your list:

```bash
sudo ubuntu-drivers list --gpgpu
sudo ubuntu-drivers install --gpgpu nvidia:535-server
```

Running `sudo ubuntu-drivers install --gpgpu` without a name installs the recommended branch instead. Then install the utilities with `nvidia-smi`; their number must match the driver branch:

```bash
sudo apt install nvidia-utils-535-server
sudo reboot
```
**Debian 13**

Enable the `contrib`, `non-free` and `non-free-firmware` components for `trixie` and `trixie-security` in your APT sources, then:

```bash
sudo apt update
sudo apt install linux-headers-amd64
sudo apt install nvidia-kernel-dkms nvidia-driver
sudo reboot
```

Debian's wiki lists which GPU generations each driver series supports and the state of the open-module flavour (`nvidia-open-kernel-dkms`); check it for your GPU. For the newest GPU generations, use NVIDIA's repository instead (next tab).
**NVIDIA repository**

NVIDIA's own repository carries the newest drivers for Ubuntu and Debian. Use the distribution name `ubuntu2404`, `ubuntu2604`, `debian12` or `debian13`:

```bash
sudo apt install linux-headers-$(uname -r)
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb
sudo apt update
sudo apt install nvidia-open
sudo reboot
```

For a headless compute server, NVIDIA also documents the smaller set `sudo apt -V install libnvidia-compute nvidia-dkms-open`. On Debian, enable the `contrib` component first.

Install the driver from **one** source only. Mixing Ubuntu's driver packages with NVIDIA's repository causes version conflicts.

**Verify:**

```bash
nvidia-smi
```

The table shows your GPU, the driver version and the highest CUDA version the driver supports.

## Step 2: Install the CUDA toolkit (only to compile CUDA code)

The toolkit comes from NVIDIA's repository. If you have not added the keyring in Step 1, add it now (same `wget` and `dpkg -i` lines), then:

```bash
sudo apt update
sudo apt install cuda-toolkit
```

The toolkit does **not** install a driver. Add its compiler to your path; the folder name carries the version, so check it first:

```bash
ls -d /usr/local/cuda*
echo 'export PATH=/usr/local/cuda/bin:$PATH' >> ~/.bashrc
source ~/.bashrc
```

If `/usr/local/cuda` does not exist, use the versioned folder that `ls` printed instead.

**Verify:** `nvcc --version` prints the CUDA compiler version.

## Step 3: Keep the driver updated

- Ubuntu driver packages update with `sudo apt update && sudo apt upgrade`. Reboot after a driver update.
- With NVIDIA's repository, NVIDIA documents version-pinning packages (`nvidia-driver-pinning-…`) if you need to stay on one branch.

## Troubleshooting

**`Failed to initialize NVML: Driver/library version mismatch`.** The driver was updated but the old kernel module is still loaded. Reboot.

**`NVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driver`.** The module did not load. On Secure Boot systems with DKMS drivers, enroll the MOK key or use Ubuntu's signed drivers; also check `dkms status` and `dmesg | grep -i nvidia`. More in [GPU troubleshooting](/guides/gpu-troubleshooting).

**`apt` reports conflicts between NVIDIA packages.** Drivers from two sources are mixed. List them with `dpkg -l | grep -i nvidia`, remove the packages of one source with `sudo apt purge` and their names, reboot, and keep one source.

**`nvcc: command not found`.** The toolkit is installed but not on your `PATH`; see Step 2.

## Next steps

- GPU containers: [GPUs in Docker containers](/guides/gpu-in-docker-containers).
- Serve a model: [install vLLM](/guides/install-vllm) or [install Ollama](/guides/install-ollama).

## Frequently asked questions

### Do I need the CUDA toolkit to run AI applications?

Usually not. Most applications and container images bring their own CUDA libraries and need only the driver. Install the toolkit when you compile CUDA code yourself.

### Which driver version should I install?

The one ubuntu-drivers recommends for compute, or the current branch from NVIDIA's repository. Check the requirements of your software: newer CUDA versions need newer drivers.

### Does this apply to a shared vGPU?

No. A shared vGPU uses the vGPU guest driver that matches the host's vGPU software; our images ship with it. Do not install the driver from this guide over it. To reinstall it, use the version listed in your service details or provided by HyperDC support.

### Does Secure Boot matter?

On systems with Secure Boot on, kernel modules must be signed. Ubuntu's ubuntu-drivers installs pre-built signed drivers; drivers built with DKMS need a machine owner key (MOK) enrolled.

---

Source: <https://hyperdc.com/guides/products/install-nvidia-driver-cuda>\
Updated: 2026-10-16