AMD and Intel unify PyTorch CPU inference under oneDNN library
The chipmakers are merging their separate optimization efforts into a single open-source library to streamline AI development for x86 processors.
Key points
- AMD is moving its ZenDNN optimizations into the main oneDNN library.
- The collaboration was recommended by the joint x86 Ecosystem Advisory Group.
- AMD becomes a steering member of the UXL Foundation with annual dues.
AMD and Intel are combining their separate CPU optimization efforts for the PyTorch machine-learning framework into a single, unified code path. This move consolidates performance tuning under the oneDNN library, which is published by the UXL Foundation and primarily driven by Intel, according to Tom's Hardware.
The integration means that AMD’s previously separate downstream version, known as ZenDNN, will now be incorporated directly into the open-source oneDNN project. This change allows developers to use a single backend for CPU inference that automatically applies the correct optimizations for either AMD or Intel hardware.
Context and Background
The decision to unify these efforts was recommended by the x86 Ecosystem Advisory Group, a joint initiative established in 2024 to foster collaboration between the two major chipmakers. The companies have previously worked together through this group, including efforts related to the ACE extensions for x86 announced earlier this year.
Intel originally developed the oneDNN library before contributing it to the UXL Foundation. Intel was among the founding members of the foundation, which also includes Arm, Qualcomm, and Google Cloud. The library is designed to detect hardware features at runtime to generate optimized kernels for the specific platform.
Who Is Affected
Developers using PyTorch for AI inference on CPUs are the primary beneficiaries of this change. The oneDNN backend was added to PyTorch in late 2023 and has already demonstrated performance improvements, including a 1.24x speedup for Float32 data types and a 2.21x improvement for BFloat16 data types.
While the news focuses on x86 processors, the oneDNN library is cross-platform and also supports Arm and RISC-V architectures. By simplifying the development environment, the move ensures that a single backend provides proper optimizations regardless of the underlying hardware being used.
What should you do?
AMD has also announced that it is joining the UXL Foundation as a "steering member." This role allows the company to influence the future direction of the foundation and the library. In exchange for this membership, AMD will contribute $20,000 per year to the organization.
What to do and how to stay safe: AMD
- Monitor the oneDNN GitHub repository for updates on the integration of AMD’s ZenDNN optimizations.
- Test PyTorch CPU inference workloads on both Intel and AMD hardware to verify performance gains.
- Review documentation for the x86 Ecosystem Advisory Group to understand future collaborative initiatives.
- Ensure development environments are updated to utilize the latest oneDNN backend versions.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Why are AMD and Intel merging their PyTorch optimizations?
The x86 Ecosystem Advisory Group recommended the move to simplify development and provide a single optimized backend for both chipmakers' CPUs.
What is ZenDNN?
ZenDNN was AMD’s downstream version of the oneDNN library, which is now being merged into the main open-source oneDNN project.
Does this affect non-x86 processors?
Yes, the oneDNN library is cross-platform and also supports Arm and RISC-V architectures, though the current news focuses on x86 collaboration.
