Intel's oneDNN 3.5 Begins Optimizing For Xe2, More Xeon 6 Tuning

Intel's oneDNN 3.5 has been released as this Deep Neural Network Library for the oneAPI specification and now part of the UXL Foundation. With oneDNN 3.5 comes more performance optimizations for existing and upcoming Intel hardware.

The oneDNN 3.5 release has improved performance for 4th Gen Xeon Scalable "Sapphire Rapids" processors and improving the performance for Xeon 6 with the recently launched SIerra Forest and the upcoming Granite Rapids CPUs. The oneDNN 3.5 release also has common tuning for enhancing the performance of the group normalization primitive, better MATMUL primitive performance, improving the performance of various subgraphs with the Graph API, and other tuning.

There is also GPU tuning with enhancements for upcoming Xe2 hardware in Lunar Lake and Battlemage. Plus there are more optimizations for the Intel Data Center GPU Max Series, better Intel Arc Graphics DG2/Alchemist performance, and other enhancements.

This library for building deep learning / AI software has also seen a number of API additions, OpenCL runtime support for its Graph API, experimental micro-kernel API for Intel processors, and FP64 MATMUL support for Intel GPUs.

Downloads and more details on the new oneDNN 3.5 release via GitHub. Fresh oneDNN benchmarks soon on Phoronix.