libdnn icon indicating copy to clipboard operation
libdnn copied to clipboard

Published 20 hours ago • verified publisher icon poweic

Switch to cuDNN

Open supergrover opened this issue 10 years ago • 3 comments

I noticed you wanted to switch to cuDNN.

I would like to keep the old CUDA code as well. For example my GPU doesn't allow for cuDNN (it's too old). We could test if the GPU supports cuDNN (I believe cuDNN is supported from compute capability >= 3 or so).

supergrover avatar Feb 01 '15 13:02 supergrover

I didn't realize that cuDNN need Compute Capability >= 3.0 Sure, we should keep the old one.

poweic avatar Feb 01 '15 13:02 poweic

Is there any documentation that tells what compute compability you need to run cuDNN v1,v2,v3?

mrgloom avatar Sep 18 '15 09:09 mrgloom

I have cuDNN v1 user guide. In CUDNN_Library.pdf, "2.4. Requirements cuDNN supports NVIDIA GPUs of compute capability 3.0 and higher and requires an NVIDIA Driver compatible with CUDA Toolkit 6.5."

To get this PDF, you have to login to developer.nvidia.com and go to the download page to see this PDF.

poweic avatar Sep 19 '15 15:09 poweic