AMD and Intel have agreed to unify their PyTorch optimization efforts within the open-source deep learning library oneDNN. This initiative is managed under the Unified Acceleration Foundation (UXL), which focuses on promoting open computing acceleration technologies. The move aims to reduce fragmentation in the x86 software ecosystem and streamline the development and deployment of AI applications for processors from both manufacturers.
The decision follows recommendations from the x86 Ecosystem Advisory Group. AMD and Intel have designated oneDNN as the primary open-source library for hardware-accelerated AI in PyTorch on x86 chips. AMD has already begun migrating optimizations from its proprietary ZenDNN library into the oneDNN codebase and has joined the UXL Foundation steering committee.
The companies are also developing AI Compute Extensions, a set of x86 instruction set extensions designed to accelerate matrix multiplication operations, which are critical for neural network training and inference. These extensions will be integrated into PyTorch via oneDNN updates. However, the companies have not yet disclosed the status of these extensions or timelines for their arrival in commercial processors.
This collaboration focuses on software infrastructure rather than the release of new hardware. The goal is to provide a unified path for developers to access existing capabilities of AMD and Intel processors, while also creating a standardized framework to support future hardware features through PyTorch.
