cuDNN¶

red requires NVIDIA cuDNN. The installation depends on a working CUDA install — see orange's CUDA & Emergent SDK guide (the eSDK section is orange-specific and can be skipped for red).

We use cuDNN 8.9.3 with driver 525.105.17 and CUDA 12.0. For other CUDA versions, pick a matching cuDNN from the cuDNN version archives.

1. Download and extract¶

Download cudnn-linux-x86_64-8.9.3.28_cuda12-archive.tar.xz from the cuDNN archives. Extract it.

2. Copy files into your CUDA install¶

Assuming CUDA is at /usr/local/cuda:

sudo cp cudnn-*-archive/include/cudnn*.h /usr/local/cuda/include
sudo cp -P cudnn-*-archive/lib/libcudnn* /usr/local/cuda/lib64
sudo chmod a+r /usr/local/cuda/include/cudnn*.h /usr/local/cuda/lib64/libcudnn*

3. Verify installation¶

source ~/.bashrc
cat /usr/local/cuda/include/cudnn_version.h | grep CUDNN_MAJOR -A 2

Expected output:

#define CUDNN_MAJOR 8
#define CUDNN_MINOR 9
#define CUDNN_PATCHLEVEL 3
--
#define CUDNN_VERSION (CUDNN_MAJOR * 1000 + CUDNN_MINOR * 100 + CUDNN_PATCHLEVEL)

cuDNN is now available to OpenCV (rebuild OpenCV with WITH_CUDNN=ON — see orange's OpenCV guide, Build option A) and to any other framework that links against it.