AMD and Intel are combining their PyTorch CPU optimisations into oneDNN, giving AI developers fewer vendor-specific software paths to manage.
Following a recommendation from the x86 Ecosystem Advisory Group, the rivals have agreed to use oneDNN as the default open-source library for x86 acceleration within PyTorch.
AMD is integrating its ZenDNN optimisations into the shared library and has joined the Unified Acceleration Foundation, which governs oneDNN, as a steering member.
The aim is to give developers one place to obtain CPU optimisations for both companies’ processors, simplifying application development, maintenance and deployment.
Intel vice president and general manager of Data Center Strategic Programs Jeff McVeigh said the collaboration would provide a consistent approach across modern processors.
He said: “Developers shouldn’t have to think about separate software paths to get the best from an x86 CPU. Working with AMD through oneDNN gives the PyTorch community a consistent approach for modern processors and a way to bring future capabilities, including AI Compute Extensions, into the framework.”
Chipzilla and AMD are collaborating on AI Compute Extensions, or ACE, an instruction set designed to accelerate matrix multiplication.
That operation sits at the heart of many AI workloads. Future oneDNN updates will support ACE, bringing the capabilities into PyTorch through the same shared software foundation.
AMD corporate vice president of Architecture and Strategy Robert Hormuth said the work would make high-performance AI applications easier to develop.
“AMD is advancing an open software ecosystem that gives developers greater choice and makes it easier to build high-performance AI applications. By working with the x86 Ecosystem Advisory Group, we’re bringing the industry together around common technologies that simplify development across x86 processors,” he said.
AMD says it will continue optimising for its processors while contributing that work through oneDNN. A shared library therefore leaves room for hardware-specific tuning while keeping development on a common path.
The announcements provide no benchmark results or specific PyTorch release date for the combined optimisations. ACE integration remains planned for future updates.
