Cuda Version compatibility with NVIDIA RTX 4090 [UbuntuOS 20.04]

Accelerated Computing CUDA CUDA Setup and Installation
, , ,
1

Hello Nvidia team and community,

I am looking to get a new system with the below mentioned configuration. I would like to know if the RTX4090 which is based on NVIDIA Ada Lovelace Architecture will support the cuda version 11.2 and cudnn 8.1 to run the tensorflow 2.6. Or what would be the minimum version of cuda supported by the 4090 series? Because it is very unclear and not mentioned in the docs here - CUDA Compatibility :: NVIDIA Data Center GPU Driver Documentation

System Spec looking to get :

CPU : Intel Core i9-13900K
GPU : NVIDIA GeForce RTX 4090

Thanks!

2

Ever get a reply here? I have a 4090 to find out the documented cuda requirement is 12 which is not supported currently by any tensorflow release. So maybe downgrade to the highest supported and ‘see’ or wait for a tensorflow upgrade?

3

I installed tensorflow-gpu using conda. I think it comes with cuda 10.
However my system cuda version is 12.
I was able to run tensorflow1 on RTX 4090, and I think tf2 should work too. So in general you should be fine.

1 Like
4

Although in the official CUDA toolkit documentation RTX 40 series support starts with CUDA 11.8. But I found that RTX 4090 also work well under CUDA 11.3+ (currently using pytorch 1.10.0 which support cuda 11.3).
Back to the question, CUDA 11.2 may not be fully compatible with RTX 4090, but is worth to take a try.