Dedicated servers · Question & answer

How do I check NVIDIA driver and CUDA compatibility on a dedicated server?

Compare the GPU architecture, installed driver and application runtime with NVIDIA’s current compatibility guidance. The CUDA value shown by nvidia-smi describes the maximum CUDA version supported by that driver; it does not identify the installed CUDA Toolkit. Applications and containers can carry their own runtime libraries.

IntermediateLinux dedicated servers with an NVIDIA GPU and installed driver
Go to the answer

Before you start

  • The application/framework version and its documented CUDA requirements.
  • nvcc is needed only for the optional installed-toolkit check.

Instructions

1. Read the driver information

Record the driver and device model without changing either. Compare them with the vendor matrix for the exact GPU architecture and intended application. Support can vary by driver branch, toolkit release and GPU generation.

Remote Linux shell
nvidia-smi

Displays the driver’s GPU summary and supported CUDA capability. It does not install or update CUDA.

2. Identify the application’s actual toolkit or runtime

If the CUDA compiler is installed and available in PATH, inspect its version. Also check the application environment or container manifest, because its runtime may differ from the host compiler.

Remote Linux shell
nvcc --version

Reports the selected CUDA compiler toolkit version when installed; absence does not prove that a runtime-only CUDA application cannot work.

3. Resolve mismatches through a planned change

Use the application and GPU vendor’s supported combination. Consider kernel, operating-system and container requirements before planning a driver change. Validate in an appropriate environment and coordinate any service interruption; do not overwrite a working driver as a diagnostic shortcut.

Check the result

  • The device architecture, driver and actual application runtime form a supported combination.
  • A representative application operation successfully uses the intended GPU.

Troubleshooting

nvidia-smi works but the application cannot load CUDA
Inspect the application’s runtime libraries, environment, architecture and container GPU exposure. Driver visibility alone does not establish runtime compatibility.

Official references

Use the documentation for your installed version when a flag, service name or platform feature differs.