oneDNN 3.14 released with ACE support and AMD Zen optimizations
Key Takeaway
oneDNN 3.14 introduces initial AMD Zen CPU optimizations and support for AI Compute Extensions (ACE).
Summary
The UXL Foundation has released oneDNN 3.14, providing optimized deep learning operations for frameworks like PyTorch and TensorFlow. This version is notable for integrating AMD Zen optimizations directly into the upstream project, moving away from AMD's separate ZenDNN fork. It also adds support for AI Compute Extensions (ACE) and improves performance for upcoming Intel Xeon Diamond Rapids and Intel GPU D-Series hardware.
Technical improvements in oneDNN 3.14
- AMD Zen optimizations: Improved MATMUL and inner-product primitive performance on Zen processors. Requires
ONEDNN_X64_USE_ZEN=ONCMake flag and ZenDNN code. - AI Compute Extensions (ACE): Initial support for instructions intended for future AMD and Intel processors.
- Intel Xeon Diamond Rapids: Performance improvements using AVX10.2 and AMX.
- Intel GPU support: Enhanced performance for upcoming D-Series (Crescent Island) and Xe3P-LPG with Nova Lake P processors.
Why it matters
The collaboration between Intel and AMD on oneDNN upstream signals a shift toward unified software optimization for deep learning frameworks across different CPU architectures.
What to do
- Users wanting AMD Zen optimizations must build oneDNN with the ONEDNN_X64_USE_ZEN=ON CMake build option.
- Users can use the ONEDNN_ZENDNN_FETCH=ON option to attempt to automatically download ZenDNN during the build process.
Original reference: www.phoronix.com →