NVIDIA cuDNN

What NVIDIA cuDNN is

NVIDIA cuDNN is a free GPU-accelerated library of primitives for deep neural networks, listed as free under the NVIDIA Software License Agreement. It supplies tuned implementations for convolution, attention, matrix multiplication, pooling, and normalization. The Graph API models operations as nodes connected by tensor edges, with flexible fusions around matrix multiplication and convolution plus specialized attention and convolution-with-normalization patterns. NVIDIA recommends its open-source Python/C++ Frontend API, while a lower-level C Backend API is also available. cuDNN accelerates frameworks including PyTorch, JAX, Caffe2, Chainer, Keras, MATLAB, MxNet, PaddlePaddle, and TensorFlow. Installation options include pip, Docker, and package managers. The backend supports Linux and Windows, but compatible operating system, CUDA, driver, and NVIDIA hardware versions are required. Only one CUDA Toolkit version of cuDNN 9 can be installed at a time.