PyTorch developers working across AMD and Intel CPUs could have fewer separate optimization paths under a oneDNN plan the chipmakers announced Oct. 20.
The companies announced the plan Oct. 20 after the x86 Ecosystem Advisory Group recommended oneDNN. The open-source library provides optimized operations, including matrix multiplication and convolution, used in AI workloads. AMD and Intel say a common software path should make applications easier to build and maintain for their processors.
AMD will contribute optimizations from its ZenDNN project to oneDNN. It has also joined the Unified Acceleration Foundation (UXL), which oversees the library, as a steering member. The role gives AMD a place in the foundation governing the software the chipmakers have chosen for their PyTorch work.
The collaboration builds on oneDNN’s existing role in PyTorch. PyTorch’s release notes say version 2.0 introduced a default X86 quantization backend that uses both oneDNN and FBGEMM. The companies’ plan would extend their work through a library that is already part of the framework.
Stephen Watt, vice president in Red Hat’s Office of the Chief Technology Officer, a distinguished engineer and a PyTorch ambassador, said the effort should make AI workloads easier to develop and deploy across x86 processors. Developers are the intended beneficiaries of the shared approach, which addresses the work involved in supporting processors from both companies.
AMD and Intel also plan to add support for AI Compute Extensions to PyTorch through future oneDNN updates. The x86 instruction set is designed to accelerate matrix multiplication, a calculation used in AI workloads. The companies have not announced a date for those updates.
