AMD and Intel make oneDNN the default PyTorch library for x86 CPUs

AMD and Intel now back one library to speed up AI work in PyTorch on x86 processors. Developers running models on a Ryzen, EPYC, Core or Xeon CPU should no longer need a separate path for each brand.

AMD says the two companies chose oneDNN, an open-source deep learning library, as the default for x86 acceleration in PyTorch. The x86 Ecosystem Advisory Group, which the two rivals set up, recommended the move.

AMD is moving its own ZenDNN optimisations into the shared oneDNN library

Until now, AMD shipped its CPU tuning for AI frameworks as ZenDNN. Intel built oneDNN. AMD is now folding ZenDNN into oneDNN and says it will keep tuning that code for its own chips.

oneDNN is run by the Unified Acceleration Foundation, or UXL, not by Intel alone. AMD has joined UXL as a steering member, which gives it a formal say in where the library goes.

"Developers shouldn’t have to think about separate software paths to get the best from an x86 CPU," said Intel’s Jeff McVeigh.

A shared matrix instruction set called ACE will reach PyTorch in later updates

AMD and Intel are also working together on AI Compute Extensions, or ACE. It is a new x86 instruction set built to speed up matrix multiplication, the core maths behind neural networks.

ACE support will come to PyTorch through future oneDNN updates. AMD gave no date.

Developers running AI models on CPUs get one code path instead of two

Nothing changes for gamers today. The benefit is for people who build or run AI tools on the CPU, such as small local models on a laptop without a strong GPU, or inference on servers.

One library means one set of bugs to fix and one place to tune. Whether AMD chips run as well through oneDNN as they did through ZenDNN is something developers will want to measure.