AMD Joins UXL as Chipmakers Set Shared x86 CPU Path for PyTorch

AMD has joined the Unified Acceleration Foundation and will work with Intel on a shared PyTorch optimization path for x86 CPUs, aiming to simplify software support across both chipmakers.

The companies selected oneDNN as the open-source library for their x86 acceleration work in PyTorch. AMD plans to integrate its ZenDNN optimizations into the library, bringing the companies’ CPU optimization efforts into a common development path.

The plan followed a recommendation from the x86 Ecosystem Advisory Group and was announced Oct. 20. The companies intend the shared approach to make applications easier to build and maintain across AMD and Intel processors.

PyTorch already uses oneDNN, including alongside FBGEMM in its default x86 quantization backend. The planned work builds on that existing use; it does not mark oneDNN’s first appearance in PyTorch or establish that every CPU operation will move to the library.

oneDNN provides optimized AI operations such as matrix multiplication and convolution. The Unified Acceleration Foundation governs the library, and AMD’s steering membership gives the company a role in the foundation.

“AMD is advancing an open software ecosystem that gives developers greater choice and makes it easier to build high-performance AI applications,” said Robert Hormuth, AMD corporate vice president of Architecture and Strategy.

AMD and Intel also plan to support AI Compute Extensions, designed to accelerate matrix multiplication, through future oneDNN updates for PyTorch.