Intel oneDNN 3.1 Further Optimizing For Sapphire Rapids, Starts Tuning For Sierra Forest

Intel's oneAPI software engineers are closing out the quarter by releasing oneDNN 3.1 as the newest version of this neural network library that is used by the likes of PyTorch, Tensorflow, PaddlePaddle, ONNX, OpenVINO, MATLAB's Deep Learning Toolbox, Apache MXNet, and many other applications.

Intel's oneAPI Deep Neural Network Library is one of their many successful open-source projects enjoying nice industry traction to help accelerate deep learning applications on a variety of CPU architectures as well as GPUs. Back in December prior to the 4th Gen Xeon Scalable "Sapphire Rapids" launch was the unveiling of oneDNN 3.0 with the polished Sapphire Rapids support and many other changes.

Out today is oneDNN 3.1 that brings continued performance optimizations for Sapphire Rapids. The oneDNN 3.0 release already has done wonders with Sapphire Rapids CPUs with the AMX support and with this new release should now be even more performant.

In addition, oneDNN 3.1 brings initial optimizations for Xeon Scalable "Sierra Forest" CPUs that are the E-core-only Xeon CPUs with up to 144 cores per socket and releasing in the first half of 2024. While still a year from seeing Sierra Forest, thanks to Intel's great open-source software track record they are already busy working on performance optimizations for that processor.

The oneDNN 3.1 release also continues optimizing the graphics performance not only for their Data Center GPU Max Series but also Arc Graphics and the Data Center GPU Flex Series.

The oneDNN 3.1 release also brings some AArch64 enhancements, optimized MATMUL for AMD GPUs, improved pooling primitive performance for RISC-V CPUs with the RVV extension, enabling the oneDNN Graph API as a production feature, and a variety of other enhancements.

Downloads and more details on all of the oneDNN 3.1 release improvements via GitHub. I'll be trying out this new oneDNN 3.1 release soon.