Dependencies between CUDA Toolkit verison, driver version and Compute Capability

Hello everyone.

I am trying to install driver for NVIDIA graphics card and CUDA Toolkit. I have GeForce RTX 2060 (Notebooks). I found out there is dependency between Compute Capability (for my GPU it is 7.5) and CUDA Toolkit version. I have seen that information on follow page: https://en.wikipedia.org/wiki/CUDA (GPUs supported). I didn’t find any official information for this. Something that I found is values of Compute Capability for some GPU - https://developer.nvidia.com/cuda-gpus.

On this link, there is information about drivers which has to be install for appropriate CUDA Toolkit version(https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html) - CUDA driver, Table 1. Could you help me to find some official information for dependency between CUDA Toolkit version and Compute Capability? I suppose I can install CUDA Toolkit 10.1 and it’s going to work correctly on RTX 2060, but I need information about it from NVIDIA official documentation.

Thank you in advance!

Best regards,
Nemanja

From official docs, table 1:

https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html

Supported CUDA compute capability by CUDA toolkit version:

https://stackoverflow.com/questions/28932864/cuda-compute-capability-requirements/28933055#28933055

Yes, that last link is not official documentation. It is accurate.

If you would like to see a change to CUDA documentation, please file a bug report using the instructions linked in a sticky post at the top CUDA programming sub-forum.

Also note that specific GPUs (e.g. RTX2080Super, for example) may require the “latest” driver version at the time that that GPU was released to the public. So even though CUDA 10.0 “supports” the Compute capbility 7.5 GPUs including RTX Super, it may still be necessary to use a newer driver to support that GPU. I don’t have a table of minimum driver versions needed for specific GPUs. The general advice is simply to install the latest version available for your GPU. That driver version will support all previously released CUDA toolkits as of the time of that driver release.

I have laptop with rtx2060 card, Windows 10 and nvidia toolkit 10.1 installed All “simple” samples run OK but during build of tensor core sample I got missed nvcuda.dll errror missed error for x64 even after Geforce driver update to the most recent version 436.30-notebook-win10-64bit-international-dch-whql.exe. From other hand I found half dozen nvcuda.dll, nvcuda32.dll and nvcuda64.dll files in different folders with different sizes. Which one is correct DCH driver and how to make it exclusively used by all solutions?

Hi Robert.

Thank you for the information.

I understand what you want to accent. Anyhow, the best option is install the latest driver. But for writing publication I need to add reference about CUDA Toolkit version and explain which version we can use on appropriate architecture. It should be the official information.

Best regards,
Nemanja

Well, for your GPU, you mention compute capability 7.5 That is Turing.

The CUDA 10.0 release notes mention that CUDA 10.0 adds support for the Turing architecture.

https://docs.nvidia.com/cuda/archive/10.0/cuda-toolkit-release-notes/index.html#cuda-general-new-features