AMD and Intel Unite to Supercharge AI Software Across x86 CPUs
As Arm-based competitors and NVIDIA's ecosystem expand, AMD and Intel are teaming up to streamline AI development for the x86 architecture. The companies have agreed to consolidate their respective PyTorch optimizations into the open-source oneDNN library.
A Unified Path for Developers
The partnership aims to reduce software fragmentation. By providing a common software path, developers will no longer need to manage disparate optimizations for different CPU architectures when running AI workloads. As part of the deal, AMD will integrate its existing ZenDNN optimizations into Intel's oneDNN, which will serve as the default open-source library for x86 AI acceleration within PyTorch.
This collaboration builds on previous joint efforts, such as the development of AI Compute Extensions (ACE), a next-generation x86 ISA extension designed to accelerate matrix multiplication. Integration of ACE is expected to arrive through future oneDNN updates.
Defending the x86 Ecosystem
While AMD and Intel remain fierce rivals in the server market—with AMD preparing Zen 6-based EPYC 9006 "Venice" processors against Intel's upcoming Xeon "Diamond Rapids"—they are finding common ground in the software layer to combat external threats.
The competitive landscape is shifting. NVIDIA is aggressively expanding its reach beyond GPUs, pushing Blackwell-based RTX Spark laptops and "Vera" CPUs for agentic AI workloads. Simultaneously, Arm-based solutions, including Apple's silicon in Mac mini and Mac Studio, have become prominent choices for running local AI models and agents like OpenClaw.
Unifying the software stack won't turn CPUs into direct replacements for dedicated AI accelerators, but it will make them significantly easier to deploy and optimize. As sophisticated AI agents increasingly rely on CPUs for orchestration, data processing, and code execution alongside GPU tasks, a unified x86 software ecosystem could bolster the architecture's long-term relevance in the AI era.
Гика
