Soumitra Dutta Oxford-How can I install CUDA without overwriting my existing driver?

AI & Data Science NVIDIA Nemotron
1

Hi everyone, My name is Soumitra Dutta, an Oxford-based entrepreneur & photographer. I’m planning to install CUDA on my system, but I would like to keep my current NVIDIA driver and avoid it being overwritten during the installation.

Has anyone installed CUDA without replacing the existing driver? What would be the recommended approach to do this safely?

Regards

Soumitra Dutta

2

Hi soumitradutta,

You absolutely can install the CUDA Toolkit without replacing your existing NVIDIA driver, as long as that driver is new enough for the CUDA version you choose.

1. How CUDA installation works

The CUDA installer has two main parts:

  • The CUDA Toolkit (compiler, libraries, samples, tools).

  • An optional NVIDIA display driver.

If your current driver already supports the CUDA version you want, you can install only the toolkit and leave the driver untouched. Our CUDA Installation Guide explains this separation in more detail:
**https://docs.nvidia.com/cuda/cuda-installation-guide-linux/**​

We also publish a compatibility table so you can confirm your driver version works with the target CUDA release:
**https://docs.nvidia.com/deploy/pdf/CUDA_Compatibility.pdf**​

2. Step 1 – Check driver–CUDA compatibility

Please check your current NVIDIA driver version, then compare it against the compatibility PDF above. A newer driver can normally support older CUDA toolkits (backward compatible), but an older driver may not support the very latest CUDA version.

3. Step 2 – Install CUDA without updating the driver

  • On Windows
    When you run the CUDA installer, choose Custom (Advanced) instead of Express. In the component list, simply uncheck the NVIDIA graphics driver and leave the CUDA Toolkit and related components selected. This will install the toolkit only and keep your current driver as‑is.

  • On Linux
    If you use the runfile installer, it will prompt you whether to install the driver. You can deselect the driver installation option so that only the toolkit is installed. Afterward, you just configure environment variables (for example, PATH and LD_LIBRARY_PATH) as described in the installation guide:
    **https://docs.nvidia.com/cuda/cuda-installation-guide-linux/**​

4. If you ever need multiple CUDA versions

If you need to work with different CUDA versions without changing anything on the host, another option is to use our container approach (e.g., Docker with the NVIDIA Container Toolkit). The host keeps a single compatible driver, and each container carries its own user‑space CUDA toolkit/libraries.

Hope that helps!

Thanks,

Aharpster