oneDNN 3.14 brings AMD Zen optimizations into the upstream library

The oneDNN 3.14 release starts integrating AMD Zen optimizations into the upstream library used by PyTorch, TensorFlow, and OpenVINO, reducing reliance on AMD’s separate ZenDNN fork.

The new MATMUL and inner-product optimizations are opt-in at build time and currently require ZenDNN code in the tree or fetched during the build. The release also adds initial support for AI Compute Extensions planned for future AMD and Intel processors, plus performance work for upcoming Intel CPUs and GPUs. Phoronix’s discussion page contained only the article stub, so there were no reader reactions to report.