AMD and Intel Bring CPU AI Optimizations Together in oneDNN

AMD and Intel are combining their CPU optimization work for PyTorch in oneDNN, bringing support for both companies’ processors into a shared library. The integration is intended to simplify CPU-based artificial intelligence inference by providing a single backend with tuning for hardware from either chipmaker.

The change brings AMD’s ZenDNN, which the company has maintained as a downstream version of oneDNN, into the open-source oneDNN project. Published by the UXL Foundation and driven by Intel, oneDNN will incorporate the AMD work rather than leaving those optimizations in a separate downstream library. For developers using PyTorch across different systems, the goal is to make the appropriate hardware optimizations available through the same backend.

AMD also announced that it is joining the UXL Foundation as a steering member. That membership gives the company influence over the foundation’s future in return for an annual contribution of $20,000. The move accompanies AMD’s decision to bring its CPU optimization efforts into the shared project.

According to AMD, the recommendation to consolidate the work under oneDNN came from the x86 Ecosystem Advisory Group. Established in 2024, the group is a joint initiative centered primarily on Intel and AMD. The companies have already worked together through it on other efforts, including the ACE extensions for x86 that became known earlier in 2026.

The oneDNN backend accelerates CPU inference in frameworks such as PyTorch, with optimizations designed for recent processors. When it was introduced as a PyTorch backend in late 2023, reported performance improvements were 1.24 times for the Float32 data type and 2.21 times for BFloat16. Those figures relate to the earlier backend introduction, rather than the newly announced integration of AMD’s work.

The library’s GitHub documentation describes a system that identifies hardware capabilities at runtime and generates kernels optimized for the platform. It also supports low-precision data types and includes optimized instructions for frequently used AI and machine-learning operations, including matrix multiplication and convolution.

Although the collaboration focuses on the two x86 processor companies, oneDNN also supports Arm and RISC-V architectures. Its scope therefore extends beyond the hardware at the center of the AMD and Intel announcement.

Intel originally developed oneDNN from earlier deep-learning libraries before the project became part of the UXL Foundation. Intel was a founding member of the foundation alongside Arm, Qualcomm, Google Cloud and others. AMD’s participation adds its optimization work and steering membership to that existing organization.