AMD will move its ZenDNN optimisations into oneDNN, the open source deep learning library that Intel maintains, under an arrangement the two companies described this week. The result is a single software path for running PyTorch workloads on x86 CPUs.
The decision follows a recommendation from the x86 Ecosystem Advisory Group, a body the two chipmakers joined to coordinate on the instruction set they both license. AMD is also taking a steering seat at the UXL Foundation, which governs oneDNN.
The practical effect is less fragmentation. Until now a developer tuning inference on AMD silicon and Intel silicon worked through separate libraries with separate tuning paths. Filing both into one repo means a model optimised for one vendor’s cores should carry over to the other’s without a rewrite.
The two are also working on AI Compute Extensions, a new instruction set aimed at matrix multiplication. Support is planned to arrive in PyTorch through later oneDNN releases.
There is a commercial logic underneath the cooperation. Inference on CPUs is the cheapest place to run small models and the overflow tier for large ones, and both vendors lose share to accelerators every time a customer decides the software work is not worth it. A shared library lowers that barrier for the whole x86 installed base.
It also puts the pair in a more coherent position against alternatives. Arm has spent two years building out its own server software stack, and RISC-V vendors are now shipping parts that boot standard Linux distributions out of the box. A fragmented x86 toolchain was the weakest part of that contest.
