GPU Support in Virtual Servers

Overview

Catalyst Cloud provides several options for deloying instances with NVIDIA GPU acceleration that can be utilised by selecting the desired GPU flavor when creating the instances.

All GPU instances require suitable drivers to be installed in the operating system before the GPU can be utilised. This applies to both operating system images you upload, and to images provided by Catalyst Cloud.

All GPU instances use direct pass-through of the entire GPU hardware, so any driver that is compatible with the GPU hardware can be installed. Catalyst Cloud recommends using drivers supplied by the operating system vendor or downloaded directly from NVIDIA.

These steps do not apply to Kubernetes worker nodes. For more information on using GPU with Kubernetes, please refer to the cluster GPU acceleration documentation.

Creating a GPU-backed Instance

To create a GPU-backed instance, simply create an instance using a GPU-backed flavor.

The list of GPU-backed types and which GPUs they provide is listed under the Accelerated Types section of the Instance types documentation.

Once the instance is created, suitable GPU drivers will need to be installed before the GPU can be used.

Setup for most GPU instance types

Minimum Requirements

The minimum requirements for GPU instances are as follows:

  • A boot/OS disk of at least 30GB (when installing CUDA support). Windows instances should use at least 50GB.

  • Compatible NVIDIA GPU driver.

The following operating systems have been tested with GPU support:

  • Ubuntu LTS 20.04 and later.

  • Rocky Linux 8 and later.

  • Windows Server 2019 and later.

All other OS images are unsupported and untested.

GPU Driver Installation

Basic Installation Process

The simplest method to install drivers is to use the packages supplied in the operating system repositories. The ‘server’ driver packages are recommended:

sudo apt install nvidia-driver-570-server

Note

NVIDIA recently began providing the option of either a proprietary or an open source driver. As of the 570 release the open source driver does not work on g4 (4x GPU) flavours on Catalyst Cloud.

Once the driver is installed, use nvidia-smi to verify that the driver is loaded and the GPU(s) detected:

$ nvidia-smi
Mon Jun 30 02:40:15 2025
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 570.133.20             Driver Version: 570.133.20     CUDA Version: 12.8     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA L40S                    Off |   00000000:04:00.0 Off |                    0 |
| N/A   29C    P0             84W /  350W |       0MiB /  46068MiB |      3%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+

If that doesn’t work, try running sudo modprobe nvidia to ensure the NVIDIA driver is loaded, or reboot the instance.

Please refer to the CUDA section for instructions on installing the CUDA toolkit.

Automated Driver Installation

For a more streamlined setup of GPU instances, the necessary GPU driver packages can be installed via user data when creating instances. This means the GPU is ready to use within a few minutes of the instance booting up without requiring additional steps.

User data example for automatically installing NVIDIA driver release 570 on Ubuntu 24.04:

#cloud-config

package_upgrade: true
packages:
  - nvidia-driver-570-server

Other versions of Ubuntu and other distributions may require a different package name. Please refer to the documentation for the specific distribution for more examples.

CUDA Support

The CUDA version supported is specific to the NVIDIA driver release version. In most cases simply download the appropriate CUDA toolkit version from NVIDIA to match the driver release used in the instance:

https://docs.nvidia.com/cuda/cuda-installation-guide-linux/

Some operating systems (e.g. Ubuntu) include CUDA packages in their repositories that can also be used instead, although they are usually older versions.

Docker Support

NVIDIA provide documentation on supporting GPU access from Docker containers here:

https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html