How to know which cuDNN version one should use?

Question

I plan to use cuDNN on Linux: how to know which cuDNN version I need? Should I always use the most recent one?

E.g. choosing the right CUDA version depends on the Nvidia driver version. I wonder if there is similar constrains for choosing the cuDNN (given that it may give some fancy error messages later on I'd prefer to know before I try).

Robert Crovella · Accepted Answer · 2016-08-18T16:28:44.170

13

You should use whichever is the latest version of cuDNN supported by your application and platform, since that will have the most bug fixes and enhancements. And yes, cuDNN versions depend on specific cuda versions. That is spelled out in the download page. (You will have to be a registered developer to access that page.)

cuDNN v5.1 has different versions for CUDA 7.5 and CUDA 8.0
cuDNN v5   has different versions for CUDA 7.5 and CUDA 8.0
cuDNN v4 and v3 both require CUDA 7.0
cuDNN v2 and v1 both require CUDA 6.5
All cuDNN versions require compute capability >= 3.0 devices

Again, all of the above information is available on the download page for cuDNN. Previous versions can be downloaded from the link at the bottom of the download page.

edited Aug 18 '16 at 16:28

answered Nov 17 '15 at 15:08

Robert Crovella

143,785
11
213
257

Have you tried cuDNN v3 on CUDA 6.5? What is the potential issue if using a lower version of CUDA? – Jon Apr 16 '16 at 05:55
According to the user guide, cuDNN v5.1 also works with CUDA Toolkit 7.0 for ARM platforms. – Franck Dernoncourt Dec 24 '16 at 14:40

Franck Dernoncourt · Answer 2 · 2016-12-24T14:52:18.473

4

In addition to CUDA version requirements, you need to ensure that your GPU has compute capability that is high enough:

All versions of cuDNN from 1.0 to 5.1 require compute capability 3.0 or higher.

Wikipedia's CUDA article has a good list of the various compute capability levels and corresponding GPUs (a similar list is available on Nvidia's compute capability webpage).

edited Dec 24 '16 at 14:52

answered Nov 17 '15 at 15:17

Franck Dernoncourt

77,520
72
342
501

I've edited the salient compute capability information into @RobertCrovella answer, for the sake of completeness – talonmies Nov 17 '15 at 19:04
Where we can check info for CuDNN v4, v5 ? – mrgloom Jul 20 '16 at 10:06
@mrgloom, I've updated the answer to reflect the new versions (they all require 3.0), but you can check the user guides of future releases [here](https://developer.nvidia.com/rdp/cudnn-download) (you'll need an account). – James Dec 24 '16 at 10:33
1

Does anyone know why Nvidia requires users to register on their website to access user guides and downloads? – Franck Dernoncourt Dec 24 '16 at 14:50

score 0 · Answer 3 · answered Jan 24 '23 at 12:06

From CUDA version 11.6 and onwards, there is no need for cuDNN version to match CUDA version. You can install whatever the latest version is available. Like I have installed cuDNN version released for CUDA 11.8 with CUDA 11.6. Here you can see at Nvidia's website that they have generalized 11.x for latest cuDNN downloads.

How to know which cuDNN version one should use?

3 Answers3

Linked