Intel's oneDNN Neural Network Library Prepares For Lunar Lake Xe2, Sierra Forest & GNR

Intel has published oneDNN 3.4 as the newest version of this Deep Neural Network Library that is part of their oneAPI software collection. The oneDNN library provides deep learning primitives for software like PyTorch, MXNet, ONNX Runtime, OpenVINO, MATLAB Deep Learning Toolbox, and other sotware.

The oneDNN CPU and GPU engines continue to support a variety of targets both internal and external to Intel. With the oneDNN 3.4 release there are more performance improvements for Sapphire Rapids / Emerald Rapids plus performance improvements for upcoming Xeon Scalable Sierra Forest and Granite Rapids processors. On the CPU side are also various AVX2 and AVX-512 improvements, some Intel AMX performance improvements for MATMUL, a few experimental CPU optimizations, and a lot of other ongoing tuning.

For non-Intel processors, oneDNN 3.4 adds support for building with the macOS Accelerate Library to enhance the performance on Apple Silicon.

On the graphics side, there is initial optimizations for Intel Xe2 graphics that are debuting with Lunar Lake processors. The Intel graphics support also includes better performance for the Data Center GPU Max series, improved performance for Arc Graphics (DG2/Alchemist), improved MATMUL performance on Intel GPUs for LLMs and transformer-like models, improved convolution performance relevant to Stable Diffusion, improved RNN primitive performance, and more.

The oneDNN 3.4 release also adds support for the Intel Data Center GPU Max 1550VG, opt-in deterministic mode support, an accumulation mode control, and other changes.

Downloads and more details on the oneDNN 3.4 release via GitHub. I'll have out some new oneDNN benchmarks soon.