The UXL Foundation announced today the release of oneDNN 3.14 as the latest feature update to this library providing optimized deep learning operations for software like PyTorch, TensorFlow, OpenVINO, and others. While oneDNN began as an Intel software project, making this version notable is beginning to see AMD Zen CPU optimizations land in oneDNN proper.
Earlier this week was the surprise announcement of Intel and AMD now collaborating on software optimizations for PyTorch with oneDNN. AMD has long maintained their own fork as ZenDNN while now AMD is beginning to contribute to oneDNN upstream.
For oneDNN 3.14 there is better performance of MATMUL and inner-product primitives on AMD Zen based processors. These optimizations are opt-in and require building oneDNN with the ONEDNN_X64_USE_ZEN=ON CMake build option. This AMD Zen work does require the ZenDNN code at least for now being within the oneDNN tree. Users will need to do that manually or there is also the ONEDNN_ZENDNN_FETCH=ON option for attempting to automatically download ZenDNN as part of the oneDNN build process.
The oneDNN 3.14 release is also exciting for having initial support for the AI Compute Extensions (ACE) instructions to be supported by future AMD and Intel processors.
This release also improves performance for upcoming Xeon Diamond Rapids processors with AVX10.2 and AMX. There are also a variety of other Diamond Rapids optimizations found in this version.
On the GPU side, oneDNN 3.14 has improved performance for the upcoming Intel GPU D-Series (Crescent Island) as well as improving performance for Xe3P-LPG with Nova Lake P processors.
Downloads and more details on the oneDNN 3.14 release via GitHub.