AMD与Intel宣布将PyTorch的x86 CPU优化统一至oneDNN，整合ZenDNN并规划ACE指令支持

AMD and Intel plan to consolidate their PyTorch optimizations for x86 processors in oneDNN, aiming to give AI developers a common path to performance improvements across both companies' CPUs and reduce the work of maintaining separate software paths.

Announced on October 20, 2026, the collaboration follows a recommendation from the x86 Ecosystem Advisory Group. The chipmakers have agreed on oneDNN as the default open-source library for x86 acceleration within PyTorch, according to their announcements.
AMD brings ZenDNN into shared library
AMD is integrating its ZenDNN optimizations into oneDNN while continuing to tune software for AMD processors. The company also said it has joined the Unified Acceleration Foundation, which governs oneDNN, as a steering member.
The stated objective is to put CPU optimizations for PyTorch in one place, making AI applications easier to build, maintain, and deploy across AMD and Intel systems. The companies framed the collaboration around a more consistent developer experience and support for existing and future x86 AI capabilities.
The oneDNN project documentation describes a library that supplies optimized implementations of operations including matrix multiplication, convolution, attention, and normalization. The library already serves as an acceleration layer for frameworks such as PyTorch and TensorFlow.
The documentation also says oneDNN identifies processor capabilities at runtime and generates specialized kernels using just-in-time compilation. Those kernels can be tailored to the available instruction set, core count, and cache sizes.
Executives emphasize developer consistency
Jeff McVeigh, Intel's vice president and general manager of Data Center Strategic Programs, said the work would give the PyTorch community "a consistent approach for modern processors" while creating a route for future capabilities to reach the framework.
Robert Hormuth, AMD's corporate vice president of Architecture and Strategy, described the collaboration as part of an effort to give developers more choice and simplify high-performance AI development across x86 processors.
Stephen Watt, a Red Hat vice president and distinguished engineer who is also a PyTorch ambassador, welcomed the agreement. He said standardizing the CPU optimizations would reduce fragmentation and make deployment across x86 architectures easier.
Chipmakers prepare for AI Compute Extensions
The collaboration also covers AI Compute Extensions, or ACE, an x86 instruction-set development intended to accelerate matrix multiplication. AMD and Intel said ACE support would reach PyTorch through future oneDNN updates.
An x86 Ecosystem Advisory Group white paper dated April 15, 2026, describes ACE as working with AVX10 vector instructions. ACE uses outer-product operations for matrix acceleration and lets chip designers tune performance to fit the available chip area.
The paper says initial software enablement includes compiler, debugger, and profiler integration, with further work planned for optimized kernels, libraries, and machine-learning runtimes.
The October announcements did not specify a PyTorch release version, an integration deadline, or benchmark results for the combined optimization effort.