cuDNN
also: CUDA Deep Neural Network library
cuDNN is NVIDIA's GPU-accelerated library for deep neural networks, providing optimized implementations of standard operations like convolution and pooling.
cuDNN (CUDA Deep Neural Network library) is a specialized library that runs on NVIDIA GPUs to accelerate deep learning workloads. It complements CUDA by providing pre-optimized, production-ready kernels for common neural network operations, making deep learning frameworks run significantly faster on GPU hardware.
Common frameworks like TensorFlow, PyTorch, and Keras integrate cuDNN to offload compute-intensive operations to the GPU. For example, a convolutional layer that would take seconds on a CPU can execute in milliseconds on a GPU with cuDNN backing it.
cuDNN is proprietary software from NVIDIA and requires both CUDA Toolkit and compatible NVIDIA GPU hardware to function. It's essential for practical deep learning on Linux systems, as most production ML workloads depend on it for performance.