cuDNN is NVIDIA's GPU-accelerated library of primitives for deep neural networks and machine learning . It is integrated into higher-level machine learning frameworks such as UC Berkeley’s Caffe software, and is now available for Jetson TK1. cuDNN Release 1 is the current stable release, and cuDNN v2 is at release candidate 3. At this time, Caffe uses cuDNN release 1. All libraries and documentation are available from the NVIDIA website, which requires a no-cost developer registration to download.
An overview of cuDNN for embedded is on the Parallel ForAll Blog.
A step-by-step procedure to install cuDNN on the Jetson is available as a GitHub gist at . After installing the cuDNN library, the Caffe
Makefile.config must be modified to use cuDNN.
# Makefile.config (only relevant cuDNN configuration items shown) USE_CUDNN := 1 INCLUDE_DIRS := /usr/local/cuda-6.5/include LIBRARY_DIRS := /usr/local/cuda-6.5/lib
After these configuration changes, the Caffe source code must be recompiled.
The NVIDIA documentation contains information on how to develop software code that utilizes cuDNN (caveat: the version 2 rc3 documentation is more thorough).