OneDNN is an open-source implementation of the oneAPI specification that provides optimized building blocks for deep learning applications. It offers primitives such as convolution, matrix multiplication, and other fundamental operations optimized for Intel and AMD CPUs, Arm processors, and various GPUs. The library serves as a performance backend for major frameworks including PyTorch, TensorFlow, ONNX Runtime, and llama.cpp, enabling developers to accelerate deep learning inference and training.