AMD and Intel will unify their artificial intelligence (AI) optimizations for x86 processors in PyTorch around oneDNN, after agreeing to make that open-source library the default option for accelerating the framework on CPUs from both manufacturers. The two companies announced the move on October 20, 2026. AMD will fold its ZenDNN library’s optimizations into oneDNN and has joined the steering committee of the UXL Foundation, the organization that governs the project.
The AMD-Intel oneDNN and PyTorch agreement in 30 seconds
- oneDNN becomes the reference library for accelerating PyTorch on AMD and Intel x86 CPUs.
- AMD will bring ZenDNN’s optimizations into oneDNN and keep tuning them for its processors.
- AMD joins the UXL Foundation’s steering committee.
- ACE instructions for matrix multiplication will reach PyTorch through future oneDNN releases.
- The announcements give no dates, PyTorch versions, or performance figures.
The decision follows a recommendation from the x86 Ecosystem Advisory Group (EAG), the group AMD and Intel created in 2024 to coordinate the evolution of the x86 architecture. Until now, getting the most out of PyTorch on a CPU meant different paths depending on the manufacturer: oneDNN, backed by Intel, and ZenDNN, AMD’s alternative for its EPYC and Ryzen chips. The stated goal is a single software path with less fragmentation.
What changes for PyTorch users on CPU
oneDNN is a library of primitives for neural networks (convolutions, matrix multiplications, normalizations) that PyTorch already uses to accelerate inference and training on processors. It originated at Intel and is now part of the ecosystem governed by the UXL Foundation, an open consortium seeking multi-platform alternatives to proprietary acceleration software stacks.
With the agreement, AMD’s specific optimizations will now live in the same project as Intel’s. In practice, a developer or platform team wouldn’t have to install and maintain a separate stack depending on whether a server runs a Xeon or an EPYC chip, which matters in data centers and clouds that mix both manufacturers. AMD says it will keep optimizing for its processors within oneDNN.
What the press releases don’t say is when all this will arrive. Neither Intel nor AMD give PyTorch or oneDNN version numbers, a ZenDNN integration timeline, or performance data. They also don’t explain what will happen to ZenDNN as a standalone project. For now, it’s a technical-direction commitment, not a change that can already be tested.
ACE, the piece coming next
The announcement also points to AI Compute Extensions (ACE), a new generation of x86 instructions designed to accelerate matrix multiplication, the core operation behind many AI workloads. According to the companies, ACE support will reach PyTorch through future oneDNN updates. cloudnews.tech previously explained what ACE is and how it aims to accelerate AI from the CPU, one of the extensions the EAG presented at its first anniversary alongside FRED, AVX10, and ChkTag.
Having a single library makes it easier for PyTorch to take advantage of ACE-equipped processors from both manufacturers once they arrive on the market. The agreement also comes as Arm gains ground in the data center. Competition between AMD and Intel continues, as shown by AMD’s advance in x86 market share, but both have an interest in keeping x86 attractive to AI developers.
What Intel, AMD, and Red Hat say
Jeff McVeigh, Intel’s vice president and general manager of data center strategic programs, sums up the approach: “Developers shouldn’t have to think about separate software paths to get the most out of an x86 CPU.” He adds that working with AMD through oneDNN gives the PyTorch community “a consistent approach to modern processors” and a way to bring future capabilities like ACE to the framework.
Robert Hormuth, AMD’s corporate vice president of Architecture and Strategy, says the company champions “an open software ecosystem that gives developers more choice” and that, through the EAG, they’re bringing the industry together “around shared technologies.” Stephen Watt, vice president in Red Hat’s Office of the CTO and a PyTorch ambassador, calls the standardization of x86 optimizations in oneDNN “a big step forward” toward ending ecosystem fragmentation.
FAQ
What is oneDNN?
It’s an open-source library of neural-network primitives that accelerates PyTorch on CPUs. It’s governed by the UXL Foundation and will become the default option for x86 on both Intel and AMD processors.
What happens to AMD’s ZenDNN?
AMD will fold its ZenDNN optimizations into oneDNN and keep optimizing for its processors. The companies haven’t explained the future of ZenDNN as a separate project.
When will the changes reach PyTorch?
No timeline or specific PyTorch or oneDNN versions have been published.
What are AI Compute Extensions (ACE)?
They’re a new generation of x86 instructions to accelerate matrix multiplication. Support for them will reach PyTorch through future oneDNN updates.
Sources:
