AMD has joined the UXL Foundation as a steering member. The move is designed to push forward work on integrating the oneDNN library with PyTorch, one of the most widely used frameworks for building deep learning models.
AMD and Intel, who have competed for decades over the CPU market, have agreed to make oneDNN the default open-source library for x86 acceleration in PyTorch.
What AMD and Intel actually agreed to
AMD’s new role became official on October 20, 2026. It marks a step up from the company’s earlier status as a general contributor to the foundation.
The headline commitment is the shared default. When PyTorch runs deep learning workloads on x86 CPUs, oneDNN is set to be the standard library handling the heavy math underneath.
The stated goal is to reduce fragmentation across AI software stacks for x86 CPUs. Today, a developer optimizing for one chip vendor can end up maintaining separate code paths for another. A shared default could trim that duplicated effort.
AMD continues to build out its own ROCm software ecosystem and to maintain its PyTorch partnerships independently of this arrangement.
Where oneDNN and the UXL Foundation fit in
The UXL Foundation was set up in September 2023. It operates under the Joint Development Foundation, which sits within the Linux Foundation.
oneDNN is one of the foundation’s core projects. It offers a unified API for high-performance deep learning primitives across x86, Arm and RISC-V CPUs, plus select GPUs.
PyTorch, TensorFlow, ONNX Runtime and OpenVINO all use oneDNN, which means a change in how oneDNN works can ripple through a large share of the AI tooling landscape.
oneDNN includes experimental support for AMD GPUs via SYCL HIP and ROCm.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.