oneDNN - GitHub README

oneDNN (oneAPI Deep Neural Network Library) is an open-source cross-platform performance library of basic building blocks for deep learning. Part of UXL Foundation. Optimized for Intel 64/AMD64, Arm AArch64, and Intel Graphics. Experimental support for NVIDIA GPU, AMD GPU, OpenPOWER, IBM z, RISC-V. Powers PyTorch, TensorFlow, OpenVINO, ONNX Runtime, llama.cpp. oneDNN Graph API provides a high-level API for fusing multiple operations into a single optimized graph. oneDNN is the default GEMM library for llama.cpp SYCL backend since 2024.8. oneDNN removed from Intel Deep Learning Essentials package in 2026.0; distributed separately since then. oneAPI Toolkit includes oneDNN until 2027.0.