Intel engineers have released the oneAPI Deep Neural Network Library "oneDNN" version 3.6 release that serves as the building blocks for deep learning software like ONNX Runtime, OpenVINO, Apache MXNet, Apache SIGNA, and optionally by PyTorch and TensorFlow with Intel's extensions.

Intel oneDNN 3.6 continues enhancing support not only for Intel's CPU and GPU products but also for other CPU architectures/vendors. The oneDNN project is technically now part of the Unified Acceleration Foundation (UXL) but in the past few years we've seen a lot of work enhancing oneDNN support for Arm and POWER processors as well as AMD and NVIDIA GPUs.

With the oneDNN 3.6 release there are performance improvements for older Sapphire Rapids Xeon processors as well as the newly-launched Xeon 6 Granite Rapids processors. There is also performance improvements for a variety of formats and different primitives.

The oneDNN 3.6 release also enhances Intel GPU support with better performance for the Intel Data Center GPU Max Series and initial support for Xe2 graphics both within Core Ultra Lunar Lake SoCs plus the upcoming Battlemage graphics processors. There is also initial optimizations for Arrow Lake H hardware.

Rounding out oneDNN 3.6 are some AArch64 processor optimizations, generic CPU support via portable SYCL kernels, enhanced NVIDIA and AMD GPU support, and experimental additions to the sparse and micro-kernel APIs.

Downloads and more details on the oneDNN 3.6 software release via [1]GitHub .



[1] https://github.com/oneapi-src/oneDNN/releases/tag/v3.6