AMD and Intel have announced a major collaboration aimed at simplifying the development process for artificial intelligence applications running on x86 central processing units (CPUs). The two semiconductor giants disclosed the partnership on October 20 local time, confirming that they will unify their respective x86 CPU optimization efforts within the PyTorch framework, specifically by integrating them into the oneAPI Deep Neural Network Library (oneDNN).
Reducing Fragmentation in the AI Ecosystem
The core objective of this joint initiative is to mitigate fragmentation within the AI software ecosystem, thereby making advanced optimization technologies more universally accessible. By consolidating these efforts, the partnership aims to enable developers to build, manage, and deploy complex AI applications with greater ease across hardware manufactured by both AMD and Intel.
Key Software Components and Technologies
The collaboration centers around two critical software tools: PyTorch and oneDNN.
PyTorch is a widely utilized collection of Python-based libraries and tools specifically designed for deep learning. It assists developers in creating and implementing AI models quickly and efficiently.
OneDNN, which stands for oneAPI Deep Neural Network Library, functions as an open-source acceleration library. It acts as a dedicated acceleration engine for x86 CPUs within PyTorch, helping to boost hardware performance and improve overall computational efficiency. Although oneDNN was initially created by Intel, it is currently maintained by the UXL Foundation under the umbrella of the Linux Foundation.
Introducing AI Compute Extensions (ACE)
Beyond the integration effort, AMD and Intel are also working together on a new architectural enhancement called AI Compute Extensions (ACE). ACE represents a next-generation extension to the x86 instruction set architecture (ISA). This technology is specifically engineered to accelerate matrix multiplication, which is recognized as a fundamental operation in the field of AI computing. The development team plans to integrate ACE into PyTorch through future updates to oneDNN.
