The Pulse
AMD and Intel Put PyTorch CPU Optimizations in oneDNN
AMD and Intel will combine their PyTorch CPU optimization work in the open-source oneDNN library. AMD plans to integrate ZenDNN optimizations, giving developers a common software path across AMD and Intel x86 processors.
AI.info Team · Written with AI · fact-checked ·
AMD plans to fold its ZenDNN optimizations into oneDNN, the open-source library Intel and AMD have selected as PyTorch’s default acceleration path for x86 CPUs. The companies announced the agreement on October 20, 2026, saying developers will be able to build and maintain PyTorch applications across both vendors’ processors using a common software foundation. The move follows a recommendation from the x86 Ecosystem Advisory Group, a collaboration between the chipmakers and other industry participants.
ZenDNN moves toward a shared library
AMD’s ZenDNN work has provided CPU optimizations for its own processors. Under the announced plan, AMD will integrate those optimizations into oneDNN rather than keep them on a separate path for PyTorch. Intel describes oneDNN as an open-source deep-learning library governed by the Unified Acceleration Foundation, or UXL Foundation.
The companies say developers will have one place to access CPU optimizations across AMD and Intel hardware. That could simplify work for teams supporting mixed x86 fleets, which otherwise may need to manage separate optimization choices or maintenance processes. AMD also joins the UXL Foundation as a Steering Member, giving it a formal role in work on the library.
“AMD is advancing an open software ecosystem that gives developers greater choice and makes it easier to build high-performance AI applications,” Robert Hormuth, corporate vice president, Architecture and Strategy, AMD said. The announcement presents the agreement as a way to reduce fragmentation, not as a claim that AMD and Intel processors will perform identically or execute the same hardware instructions.
One default does not mean one kind of CPU
Choosing a common library addresses the software layer developers use to access optimized operations. It does not erase differences between processor generations or guarantee the same speed on each system. The companies offer no benchmark results, migration instructions or completion date for the ZenDNN integration in their announcements.
Those omissions leave practical questions for developers: when the integrated optimizations will reach PyTorch users, how they will be selected for different processors, and whether existing ZenDNN deployments will need changes. For now, the commitment is to combine optimization work in oneDNN; the announcement does not say the integration is complete or specify a PyTorch release that will include it.
AI Compute Extensions remain a future step
AMD and Intel also point to their work on AI Compute Extensions, or ACE, a proposed x86 instruction-set feature designed to accelerate matrix multiplication. They say future oneDNN updates are expected to support ACE in PyTorch. The announcement does not identify processors that will support the extensions or give a schedule for their arrival.
ACE is therefore related to, but separate from, the immediate software alignment: oneDNN is the shared library plan, while ACE concerns future processor instructions and their software support. The near-term test for the agreement is whether AMD’s ZenDNN optimizations make it into oneDNN in a form developers can use across x86 systems without adding new vendor-specific maintenance work.