Intel and AMD announced on 20 October 2026 that they plan to combine their PyTorch CPU optimisations in oneDNN, giving the two rival chipmakers a shared open-source route for x86 acceleration. In its newsroom announcement, AMD said it was integrating its ZenDNN optimisations into the library and had joined the Unified Acceleration Foundation, known as UXL, as a steering member.
Key points
- The companies have selected oneDNN as the default open-source library for x86 acceleration within PyTorch.
- AMD is moving optimisations from its ZenDNN fork into oneDNN and has taken a steering role at UXL.
- Intel and AMD are also collaborating on AI Compute Extensions, an x86 instruction-set project for matrix multiplication.
ZenDNN work moves into oneDNN
oneDNN is a library for running deep-learning workloads on CPUs. Intel has long tuned it for its own processors, while AMD has maintained ZenDNN as a compatible fork optimised for Zen chips, Phoronix, a publication covering Linux hardware, reported. PyTorch and other AI frameworks have adopted oneDNN.
Intel and AMD chose oneDNN to be PyTorch’s default open-source library for x86 CPU acceleration, according to Intel’s announcement. AMD’s contribution will bring its Zen-focused optimisation work into that library. The companies say developers will be able to access CPU optimisations for both manufacturers’ processors through one software path.
Intel vice president and general manager of Data Center Strategic Programs Jeff McVeigh said the arrangement would give PyTorch developers a consistent approach across the two companies’ processors. AMD corporate vice president of Architecture and Strategy Robert Hormuth said common technologies would simplify development across x86 chips.
AMD takes a UXL steering seat
UXL governs oneDNN, the library the companies have chosen for their combined work. AMD’s steering membership places it within the foundation overseeing the project to which it is contributing ZenDNN optimisations. AMD described that membership as part of advancing the work through open collaboration.
The foundation was formed in 2023, with Intel, Samsung, Qualcomm and Arm among the vendors backing it, Phoronix reported. Projects previously associated with oneAPI, a collection of software efforts for different computing hardware, came under the consortium. AMD’s decision adds another chipmaker to that group as the companies bring their PyTorch CPU work together.
The distinction between the library and the foundation matters to the arrangement: oneDNN is where the processor-specific code is being combined, while UXL governs that open-source project. Intel says joint investment in the library is intended to reduce the separate software work involved in supporting x86 AI capabilities.
The x86 advisory group’s recommendation
The choice of oneDNN followed a recommendation from the x86 Ecosystem Advisory Group, according to Intel. AMD and Intel have also worked together through the group on instruction-set architecture matters — the instructions a processor can execute. The new collaboration applies that cooperation to software used by PyTorch applications on their CPUs.
Intel says a common route through oneDNN should make PyTorch applications easier to build, maintain and deploy across AMD and Intel systems. That is the companies’ stated objective for developers using the software, rather than a performance result from the collaboration.
AI Compute Extensions address matrix multiplication
The companies are also collaborating on AI Compute Extensions, or ACE, a proposed x86 instruction-set architecture designed to accelerate matrix multiplication. AMD describes that operation as a core part of AI workloads. The instruction-set work is separate from AMD’s integration of its existing ZenDNN optimisations into oneDNN.
Stephen Watt, Red Hat’s vice president in the Office of the Chief Technology Officer and a PyTorch ambassador, endorsed the oneDNN collaboration in Intel’s announcement. He said bringing the companies’ CPU optimisations together would make it easier for developers to build and deploy workloads consistently across x86 architectures.
.jpg){kind=link}