Does CUDA 11.8 Compatible with CUDNN 8.1.1?

AI & Data Science Deep Learning (Training & Inference) cuDNN
,
1

We are running on Windows 10, TensorFlow 2.7 (C API) to run the inference for our detection deep learning model based SSD.
The CUDA installed is 11.2.2 with CUDNN 8.1.1.33, driver 30.0.14.7247.
The typical performance of our model is 4.5 ms.

We recently upgraded the CUDA to 11.8, CUDNN 8.5.0.96, driver 31.0.15.3742
(472.12-quadro-rtx-desktop-notebook-win10-win11-64bit-international-dch-whql.exe)
to support installing onnxruntime version 1.16.3 for other project (it needs CUDA 11.8)

We noticed a degradation in time inference of ~0.5 ms.
After our investigation we found that the reason was the new CUDNN.
Thus we reinstalled back the CUDNN version 8.1.1.33 and degradation was eliminated.
We tested other CUDNN versions but this was our best combination for time performance.
At our tests we did not faced issues running: CUDA 11.8 with CUDNN 8.1.1.33 and driver 31.0.15.3742

Our question is CUDA 11.8 is formally compatible to run with CUDNN 8.1.1.33 ? does we take here any risks?

2

Hi @hamode.hg ,
Apologies for delayed response.
here is the support matrix for cuda cudnn Support Matrix — NVIDIA cuDNN 9.0.0 documentation.

I am checking with the team on this.

4

Thank you very much for the response.
Thank you for the Support Matrix you attached.
it will be great if you check specifically, that no constrains to run CUDA 11.8 with CUDNN 8.1.1.33

5

What about cuDNN v8.2.0, does it support CUDA 11.8 ?

It shows it supports “11.x”

Thanks.

6

Yes, cuDNN v8.2.0 supports CUDA 11.8.

11.x means 11.x for any x, including x=8.

7

Thank you very much for your response
Best
Hamode