AMD and Intel Unify PyTorch CPU Optimizations in oneDNN Library

Unified PyTorch Optimizations

AMD and Intel are combining their CPU optimizations for PyTorch into a single oneDNN library. This integration means that developers will have a unified code path for AI inference on both AMD and Intel chips, simplifying the development process.

Previously, AMD maintained a separate version of oneDNN called ZenDNN. This ZenDNN functionality is now being moved into the open-source oneDNN library, which is published by the UXL Foundation.

AMD Joins UXL Foundation

As part of this initiative, AMD has joined the UXL Foundation as a "steering member." This role allows AMD to influence the future direction of the UXL Foundation, requiring an annual contribution of $20,000.

Collaboration Driven by EAG

The decision to unify these efforts under oneDNN was recommended by the x86 Ecosystem Advisory Group (EAG), a joint initiative between Intel and AMD established in 2024. The two companies have previously collaborated through the EAG on projects like the ACE extensions for x86.

Intel originally developed oneDNN, building on earlier deep learning libraries, before it became part of the UXL Foundation. Intel was a founding member of the UXL Foundation, alongside other companies like Arm, Qualcomm, and Google Cloud.

Impact on Developers and Performance

The oneDNN backend is designed to accelerate CPU inference workloads for the latest Intel and AMD chips within frameworks like PyTorch. When oneDNN was added to PyTorch in late 2023, it demonstrated performance improvements of 1.24x for Float32 data types and 2.21x for BFloat16 data types.

For developers, this unification simplifies working with PyTorch across different systems, as a single backend will provide appropriate optimizations regardless of the underlying hardware. While the immediate focus is on x86, oneDNN is a cross-platform library that also supports Arm and RISC-V architectures.