On October 20, 2026, AMD and Intel announced a collaboration to consolidate x86 CPU optimizations for the AI development framework PyTorch in the open-source compute library oneDNN. AMD will integrate the optimization technology it has built up in its own ZenDNN into oneDNN, creating a shared software foundation for getting high performance out of both AMD and Intel CPUs. The aim is to spare developers from having to think about different optimization methods for each CPU vendor, making AI applications easier to develop and maintain. However, the companies have not disclosed when the integration will be completed, which PyTorch versions will be supported, or any specific performance gains. In this collaboration, the effort to integrate existing optimizations into a shared library needs to be understood separately from support for future CPU instruction sets.
AMD's ZenDNN to be integrated into oneDNN, creating a common PyTorch optimization foundation
AMD and Intel agreed to adopt oneDNN as the standard open-source library for accelerating PyTorch on x86 CPUs, following a recommendation from the x86 Ecosystem Advisory Group, in which both companies participate. According to AMD's official announcement, the company is working to integrate ZenDNN's optimization technology into oneDNN. AMD also participates as a Steering Member of the Unified Acceleration Foundation (UXL), which oversees oneDNN development, and will take part in open-source collaboration.
Through this collaboration, developers will be able to use the CPU optimizations that AMD and Intel have each developed through a common library. AMD will continue to improve performance for its own CPUs, but plans to bring those results into oneDNN rather than keeping them only in a standalone library. The two companies say that consolidating previously scattered optimizations will simplify the development, maintenance, and deployment of AI applications built on PyTorch.
In Intel's official announcement, Jeff McVeigh, who oversees data center strategic programs, stressed that developers should not need to worry about separate software paths for each vendor in order to get the most out of x86 CPUs. He said that working with AMD through oneDNN will give PyTorch developers a common optimization approach while also making it easier to support future CPU features.
That said, oneDNN already supported AMD CPUs. The official oneDNN GitHub repository explicitly lists Intel 64/AMD64 architectures as existing supported targets, so this announcement does not make oneDNN available on AMD CPUs for the first time. The key change is that AMD will bring the optimization technology it developed independently into oneDNN, and that both companies have set out a policy of pursuing PyTorch performance improvements in a shared library.
A common API, with code generated to suit each CPU
oneDNN is a library for running operations essential to AI models quickly, such as matrix multiplication, convolution, and attention. It serves as an acceleration foundation for many AI frameworks, including PyTorch, TensorFlow, and OpenVINO. Sitting between AI applications and the CPU, it is responsible for efficiently executing the processing that developers write on real hardware.
A defining feature of oneDNN is that it offers a common API while enabling optimizations tailored to the characteristics of the CPU it runs on. According to the official README, it detects CPU capabilities at runtime and uses JIT (Just-In-Time) compilation to generate computation code matched to the supported instruction sets, core count, and cache size. JIT refers to generating the machine code needed while a program is running, which makes it possible to take advantage of hardware characteristics without preparing separate code for every CPU in advance.
For example, even when a PyTorch application runs the same matrix multiplication, oneDNN can choose different processing internally depending on the instruction sets and hardware configuration the CPU supports. Optimizations suited to the CPU's vendor and generation can be applied without developers changing the API they use. This is also the significance of integrating ZenDNN's optimizations into oneDNN: by bringing acceleration techniques refined for AMD CPUs into the shared library, the benefits could become available without changes to application code.
In environments such as data centers where AMD and Intel CPUs coexist, this kind of standardization is likely to reduce software maintenance burdens. If there is less need to manage separate optimization libraries and execution paths for each CPU vendor, developers can focus more easily on the models and applications themselves. However, this is an effect expected from the collaboration, and no measurements have been published showing how much development effort or maintenance time will actually be saved.
Also, even if the way the software is used becomes common, AMD and Intel CPUs will not deliver identical performance. Differences in supported instruction sets, cache configurations, and core counts can change processing speed even when running the same AI model. This effort does not guarantee the same performance on both companies' CPUs; it makes it easier to use optimizations suited to each piece of hardware from a common software foundation.
Next-generation ACE instruction set also planned, with no implementation date
Beyond standardizing on oneDNN, AMD and Intel are also cooperating on the development of AI Compute Extensions (ACE), a next-generation instruction set architecture for x86. ACE is an instruction set designed to accelerate matrix multiplication, the core of AI processing, and the plan is to make it usable from PyTorch through future oneDNN updates. However, the integration of ZenDNN's optimizations now under way and support for ACE are separate efforts, and their implementation timing is not necessarily the same.
The current state of support can be organized into three stages: existing x86 CPU support in oneDNN, integration of ZenDNN's optimizations, and future support for ACE.
| Item | Status as of October 2026 | What developers should check |
|---|---|---|
| oneDNN support for Intel 64/AMD64 CPUs | Already supported as an existing feature | CPU architecture support is separate from the integration of new optimizations |
| ZenDNN optimizations | AMD is working on integration into oneDNN | Completion timing and the supported PyTorch and oneDNN versions have not been disclosed |
| ACE support in PyTorch | Planned through future oneDNN updates | Not a feature available today; timing and target CPUs are not yet confirmed |
This comparison is based on the October 20, 2026 announcements from AMD and Intel and the support status described in the official oneDNN repository. It is important to distinguish between the two companies agreeing to consolidate optimizations in a single library and all integration or future feature support having been completed.
The announcement also does not specify which CPU models will receive the post-integration optimizations, nor does it describe migration steps from existing PyTorch environments. When considering deployment on real systems, it will therefore be necessary to check which optimizations are included in the versions of PyTorch and oneDNN being used. As for ACE, it will be hard to build into concrete deployment plans until supported CPUs and availability timing become clear.
This AMD–Intel collaboration is significant as an effort to standardize the development environment for AI software on x86 CPUs. If the optimizations each company has pursued can be consolidated in oneDNN, developers should find it easier to build and maintain AI applications without paying excessive attention to CPU vendor differences. To judge the effect, however, it will be necessary to verify how well the integrated library performs on real AI models. Going forward, attention will focus on the progress of ZenDNN integration, support for ACE, and performance evaluations of using a common software path on both companies' CPUs.