AMD & Intel join forces on PyTorch AI optimisations

AMD and Intel have agreed to combine their x86 CPU optimisation work for PyTorch in oneDNN, creating a common software path for AI workloads across both companies' processors.

The agreement centres on oneDNN, an open-source deep learning library governed by the Unified Acceleration Foundation and used to run AI workloads on CPUs. Under the arrangement, AMD and Intel have aligned on oneDNN as the default open-source library for x86 acceleration in PyTorch.

As part of the effort, AMD is integrating its ZenDNN optimisations into oneDNN. The chipmaker has also joined the UXL Foundation as a steering member.

The work follows a recommendation from the x86 Ecosystem Advisory Group, which has been pushing for a more consistent software approach across x86 hardware as AI development on CPUs expands.

For developers, this means PyTorch CPU optimisations will be available through a single route across AMD and Intel processors. That should reduce the need to manage separate software paths when building, maintaining and deploying AI applications on x86 systems.

Common path

PyTorch is widely used by AI developers, and software fragmentation across processor platforms has been a recurring issue for teams trying to tune performance while keeping applications portable. By concentrating x86 CPU work in oneDNN, AMD and Intel are trying to simplify that process for framework users.

Robert Hormuth, Corporate Vice President of Architecture and Strategy at AMD, outlined the company's position on the collaboration.

"AMD is advancing an open software ecosystem that gives developers greater choice and makes it easier to build high-performance AI applications. By working with the x86 Ecosystem Advisory Group, we're bringing the industry together around common technologies that simplify development across x86 processors," said Hormuth.

Intel described the agreement as a way to remove added complexity for software developers targeting x86 CPUs.

"Developers shouldn't have to think about separate software paths to get the best from an x86 CPU," said Jeff McVeigh, vice president and general manager of data center strategic programs at Intel. "Working with AMD through oneDNN gives the PyTorch community a consistent approach for modern processors and a way to bring future capabilities, including AI Compute Extensions, into the framework."

Broader work

The collaboration also extends to AI Compute Extensions, or ACE, a next-generation x86 instruction set architecture for AI workloads. ACE is designed to speed up matrix multiplication, a core operation in many AI models, and will be integrated into PyTorch through future oneDNN updates.

The tie-up is notable because AMD and Intel compete directly in server and client processors, yet are choosing to work together on part of the software stack used by AI developers. The focus on shared tooling reflects the importance of software support in determining how easily developers can move workloads across different processor families.

Red Hat, which participates in the PyTorch community, backed the effort as a step towards a more unified experience for users working across x86 architectures.

"Open-source innovation thrives when the developer and operational experience is seamless and unified. AMD and Intel coming together to standardize x86 CPU optimizations through oneDNN in PyTorch is a major step forward in eliminating ecosystem fragmentation," said Stephen Watt, Vice President in the Office of the Chief Technology Officer, distinguished engineer at Red Hat, and PyTorch ambassador. "As a member of the PyTorch community, Red Hat is excited to see this level of open collaboration, which will make it significantly easier for developers to build, optimize and deploy AI workloads consistently across x86 architectures."