AMD and Intel bring PyTorch CPU optimization work into oneDNN

AMD and Intel plan to combine their x86 CPU optimization work for PyTorch in oneDNN. The move would create a shared software path across the companies’ processors.

The companies said they settled on oneDNN as PyTorch’s default open-source library for x86 acceleration following a recommendation from the x86 Ecosystem Advisory Group. AMD plans to integrate its ZenDNN optimizations into oneDNN while continuing to tune software for AMD processors. It has also joined the Unified Acceleration Foundation as a steering member to advance the work through open collaboration.

oneDNN is an open-source deep-learning library designed to help AI workloads run efficiently on CPUs. The companies say putting both vendors’ PyTorch CPU work in the library will give developers one place to obtain optimizations, instead of requiring separate software paths for AMD and Intel systems.

AMD’s current ZenDNN documentation describes zentorch as a PyTorch plug-in that applies inference optimizations for AMD EPYC CPUs through the ZenDNN library and PyTorch’s torch.compile flow. The companies have not said which ZenDNN components will move into oneDNN or whether zentorch will remain separate.

AMD and Intel said the common path should make PyTorch applications easier to build, maintain and deploy across x86 systems. They also expect it to reduce software fragmentation and speed support for current and future x86 AI capabilities. The announcements included no benchmarks or independent measurements showing lower maintenance or deployment effort.

Neither company gave an integration timetable, release milestones or a target PyTorch or oneDNN version.