oneDNN 3.14 Adds Optional ZenDNN Optimizations After AMD-Intel PyTorch Plan

The Unified Acceleration Foundation released oneDNN 3.14 on Oct. 23, adding optional AMD CPU optimizations in the first software update following AMD and Intel’s shared PyTorch development plan.

The update adds a way to use AMD’s ZenDNN optimizations for matrix multiplication and inner product, two operations used in AI workloads. oneDNN provides optimized computing operations, and the new option places AMD-specific work within the library the chipmakers selected for their joint x86 CPU effort.

The ZenDNN integration is disabled in standard builds. Developers must install ZenDNN version 6.0.0 or later separately and enable the feature. The current build path targets Linux systems with x86-64 processors, and the ZenDNN-backed operations run only on AMD CPUs.

Version 3.14 also adds initial support for AI Compute Extensions (ACE), x86 instructions AMD and Intel are developing to accelerate matrix multiplication. Developers must enable that support separately. The release does not establish whether either capability will be available by default in PyTorch.

The update follows AMD and Intel’s Oct. 20 plan for a common PyTorch CPU path. The companies chose oneDNN for their shared optimization work to make AI applications easier to build and maintain across their processors.

PyTorch already uses oneDNN. In its default x86 quantization backend, the library operates alongside FBGEMM. The companies’ plan builds on that existing use; it does not establish that every PyTorch CPU operation will move to oneDNN.

AMD joined the Unified Acceleration Foundation as a steering member as part of the collaboration. The foundation governs oneDNN. Although the library supports other processor architectures, this joint effort focuses on x86 CPUs in PyTorch.

For developers maintaining AI applications across AMD and Intel systems, the shared library provides a common place for CPU optimization work. The AMD-specific path in this release still requires a separate installation and an explicit build choice.

“Developers shouldn’t have to think about separate software paths to get the best from an x86 CPU,” said Intel Vice President and General Manager Jeff McVeigh.

Future oneDNN updates are intended to bring ACE support into PyTorch.