AMD and Intel plan shared oneDNN optimization path for PyTorch on CPUs

Where the optimization work will go

AMD’s announcement says the companies have aligned on oneDNN as the default open-source library for accelerating PyTorch on x86, the processor architecture used by their CPUs. AMD will contribute optimizations from ZenDNN, its library for inference—the execution of trained AI models—while continuing to tune software for its own processors.

Data Phoenix reports the planned integration and the companies’ goal of simplifying application development, maintenance and deployment. The intended change is a common destination for optimization work that has been maintained separately, giving developers deploying across both vendors a shared library path.

The agreement follows a recommendation from the x86 Ecosystem Advisory Group, according to AMD. CloudNews.Tech describes it as a group AMD and Intel created in 2024.

From separate tuning to a shared library

AMD’s ZenDNN repository describes a library intended for use through supported frameworks rather than independently. Its documentation describes separate integration plugins for PyTorch and TensorFlow, another AI framework.

Hardware Busters describes the move from separately maintained ZenDNN optimizations into oneDNN. It also reports that AMD previously participated in the library’s governing foundation as a general contributor.

AMD says it has now joined the Unified Acceleration Foundation, or UXL, as a steering member, giving it a place on the body governing oneDNN.

Mixed-vendor deployments are the intended beneficiaries

For teams maintaining PyTorch applications across AMD and Intel CPUs, Data Phoenix identifies the intended benefit as avoiding separate vendor optimization paths. Simpler development, maintenance and deployment are the collaboration’s stated goals.

For existing ZenDNN users, AMD’s repository still directs PyTorch users to zentorch, its integration plugin, for compatibility and installation details. That remains the documented route.

Not yet confirmed

• No integration timetable, target PyTorch or oneDNN versions, or released build containing the completed common path has been identified.
• The scope of ZenDNN components to be integrated, whether zentorch stays separate and whether ZenDNN continues as a standalone project remain unconfirmed.
• The announcement provides no measured inference speedups or demonstrated maintenance and deployment savings from this collaboration.

What to watch

• PyTorch and oneDNN release notes identifying the first integrated AMD optimizations.
• ZenDNN and zentorch documentation clarifying compatibility and the plugin’s future.
• Performance results for the integrated path across AMD and Intel processors.

Related coverage

• DevicesAmazon renames Kindle Unlimited to Kindle+, adds thousands of audiobooks
• DevicesSamsung and SK Hynix reportedly plan higher-layer NAND production at China factories
• DevicesAMD expands FSR 4.1 to RX 6000, Xbox Series X/S and more Ryzen hardware