1) cuDNN is a library of deep learning primitives for GPUs that provides highly tuned implementations of routines such as convolutions, pooling, and activation layers. 2) Version 2 of cuDNN focuses on improved performance and new features for deep learning practitioners. It supports 3D datasets and new GPUs like Tegra X1. 3) cuDNN can be easily enabled in frameworks like Caffe and Torch by making minor changes to code and is compatible with APIs for deep learning routines.