cuDNN is a program package of Windows that includes a GPU-accelerated library of primitives for deep neural networks. The software provides highly tuned implementations for standard routines such as forward and backward convolution, pooling, normalization. It also features support for activation layers.

Overview

cuDNN provides customizable data layouts, flexible dimension ordering, striding, and subregions for the 4D tensors used as inputs and outputs. Thanks to this flexibility, you can easily integrate the distribution components into any neural network implementation. Additionally, it is possible to skip the transposition steps sometimes required by GEMM-based convolutions.

Additional notes

It is necessary to mention that the cuDNN comes with a context-based API that allows for easy multithreading and optional interoperability with Nvidia CUDA streams. What is more, the provided convolution routines aim for a performance that is competitive with the fastest matrix-multiply-based implementations of such routines while utilizing significantly less memory.

Features

free to download and use;
compatible with modern Windows versions;
gives you the ability to develop deep neural networks;
provides you with customizable layouts and flexible dimension ordering;
you can integrate the functionality within existing implementations;
it is possible to perform forward and backward convolution.