The IT Home page is stamped 2026/10/21 15:08:31. The judgment first: AMD and Intel plan to put their x86 CPU optimizations into one open library. This page prints no speed comparison. More consistent performance is written as something that may follow if the work lands.
[1]
The two companies jointly announced the cooperation yesterday, October 20, on x86 CPU optimization for PyTorch. They will fold their optimization work into the open deep-learning library oneDNN, and they plan to make it the default open library for x86 acceleration inside PyTorch. AMD will integrate its ZenDNN optimizations into oneDNN and join the UXL Foundation as a steering-committee member. The committee seat is an announced arrangement. The integration and the default library are still things they will do.
[1]The page also says the two companies are advancing AI Compute Extensions, shortened to ACE. It focuses on speeding matrix multiplication, which the page calls a core operation in model training and inference. The related capability is planned to reach PyTorch later through oneDNN updates. A plan to connect it is not the same as a call that already works in PyTorch.
[1]For developers, the page says workloads may previously have needed separate tuning for AMD or Intel CPUs. The companies say one software path will lower fragmentation and cut duplicate development and maintenance. The page also says that if the work lands smoothly, PyTorch performance may feel more consistent on x86 servers, workstations, and AI PCs. Those are the companies' words. This page lists no machine, no test set, and no before-and-after speed.
[1]A note on the page calls PyTorch an open-source Python deep-learning framework, used mainly to build, train, and deploy neural networks. This cooperation sits on x86 CPU optimization inside that framework. The announcement says it will give developers one software path across the two companies' CPUs. The page does not say whether the libraries on graphics processors change with it. The shared path is a plan on the CPU side.
[1]What a reader can check comes in three layers. First, the October 20 announcement puts the optimizations into oneDNN and plans to make that the default open library for x86 acceleration in PyTorch. Second, AMD will integrate ZenDNN and join the UXL steering committee. Third, ACE focuses on matrix multiplication, and reaching PyTorch is written as a future plan. None of the three layers comes with a speed. Separate tuning is how the page describes the past. It is not a migration cost that has already been counted.
[1]