Intel Releases oneDNN 3.0 In Advance Of Sapphire Rapids
Prior oneDNN 2.x releases were already making optimizations around 4th Gen Xeon Scalable "Sapphire Rapids" processors while this oneDNN 3.0 release is said to deliver improved performance. It looks like oneDNN 3.0 is in good shape for Sapphire Rapids, especially given the timing ahead of the January announcement for Sapphire Rapids.
The oneDNN 3.0 release also introduces FP16 support and other optimizations for next-gen Xeon Scalable "Granite Rapids" processors. If past rhythms hold true for oneDNN, expect the next several oneDNN releases to further optimize this neural network library for Granite Rapids.
The oneDNN 3.0 release also improves performance for Intel Data Center GPU Max Series "Ponte Vecchio" as well as for Intel Arc Graphics DG2/Alchemist and Intel Data Center GPU Flex Series.
Intel's oneAPI effort continues to provide an excellent collection of very useful and valuable open-source software.
For quite a while now oneDNN has been supporting non-x86_64 CPUs and with oneDNN 3.0 there is a number of PowerPC and AArch64 improvements too. On the AArch64 side this Intel library for deep learning software has performance improvements around Scalable Vector Extensions (SVE) usage, and improved FP16 performance with Arm's Compute Library (ACL). On the hardware side oneDNN 3.0 is rounded out by more improvements for AMD and NVIDIA GPU usage too.
In addition to new hardware support and optimizations, oneDNN 3.0 adds a new quantization scheme, brings an experimental Graph API, adds support for Intel DPC++/C++ 2023.0, extends its thread pool API, and has other changes.
Downloads and more details on oneDNN 3.0 via GitHub. I have already started working on some new oneDNN benchmarks against the new oneDNN 3.0 release with various Intel/AMD CPUs.