Photo: Seraphfim Gallery / Pexels
The longtime rivals are merging their PyTorch optimization work into one open-source library as Arm-based competition grows
Intel and AMD have spent decades trying to beat each other. Now they are sharing a toolbox.
On October 20, 2026, the two chipmakers announced they would fold their separate x86 CPU optimization work for the PyTorch AI framework into a single open-source library called oneDNN. The aim is one software path for AI workloads, whether a developer’s machine runs Intel or AMD silicon.
What the two chipmakers actually agreed to #
Under the arrangement, AMD will bring its own ZenDNN optimizations into oneDNN. Until now, those tweaks lived in a separate stack tuned specifically for AMD processors.
AMD is also joining the UXL Foundation as a steering member. The UXL Foundation governs oneDNN, so AMD now gets a formal seat at the table where the library’s direction is decided.
The practical goal is simplifying how PyTorch-based AI applications are built and deployed on x86 hardware. Developers would no longer need to juggle one set of optimizations for Intel and another for AMD.
Intel’s Jeff McVeigh framed the move around that exact pain point. He said developers should not be stuck choosing between separate software paths just to get the best performance out of x86 CPUs.
The collaboration also promises easier access to newer chip capabilities. Features such as AI Compute Extensions, or ACE, are expected to reach developers through incremental updates to oneDNN rather than through vendor-specific workarounds.
Where the partnership came from #
This did not come out of nowhere. The project was developed under the x86 Ecosystem Advisory Group, or EAG, which Intel and AMD formed in October 2024.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The EAG has already pushed both companies toward alignment at the hardware level. It previously backed standardization of instruction set extensions, including AVX10 and ACE.
The focus is deliberately narrow. Both companies are targeting AI inference, edge computing and latency-sensitive tasks, where a CPU handling the work directly can make sense.
They are not pitching this as a replacement for GPUs. The stated intent is to complement existing GPU acceleration, not to compete with it head-on.
Why rivals are suddenly cooperating #
The motivation is fairly easy to read. Both companies have said the move strengthens the case for x86 CPUs in AI work, particularly as Arm-based architectures gain ground.
By pooling their efforts, Intel and AMD are effectively treating x86 as a shared asset worth defending.
What this means for developers, the industry and investors #
For developers, the immediate payoff is less duplicated effort. A single, jointly maintained library should mean fewer edge cases and fewer platform-specific bugs when deploying PyTorch models on x86 machines. There are open questions worth watching. Governance is one: AMD’s steering role at the UXL Foundation will show how evenly influence over oneDNN is shared in practice.
Execution is another. Merging ZenDNN into oneDNN is a meaningful engineering task, and the benefits depend on how smoothly AMD-specific optimizations perform inside a library that started life elsewhere.
The rollout of ACE support through oneDNN updates will also be a useful test. If new capabilities reach developers quickly and consistently across both vendors, the partnership will look like more than a press release.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our