NPU support for DirectML on Windows 11 is here, limited to developers as of now

Microsoft and Intel have collaborated to release a developer preview for Neural Processing Unit (NPU) acceleration within DirectML, the machine learning platform API for Windows 11. This preview supports several models on specific Windows 11 devices with Intel Core Ultra processors featuring Intel AI Boost.

In easier words, Microsoft and Intel are testing some new tools for Windows 11 that use these special chips (NPUs) to make AI tasks run smoother and quicker. Things like translating languages on the fly or unlocking your computer with your face might happen almost instantly and use less battery.

DirectML’s NPU support, as outlined in this developer preview, is targeted towards Intel Core Ultra processors with Intel AI Boost. Users interested in exploring NPU acceleration are required to have a Windows 11 device meeting this hardware specification. The necessary components include the DirectML 1.31.1 NuGet package, ONNX Runtime 1.17, and the latest NPU driver from Intel®.

However, there are certain limitations in this early release. NPU support is exclusive to Intel Core Ultra processors with Intel AI Boost, and compatibility is currently confined to a subset of machine learning models optimized for NPU acceleration.

Some models may exhibit latency, accuracy, or functionality issues. Microsoft acknowledges these limitations and is actively working towards refining compatibility and performance for a broader spectrum of models in subsequent releases.

NPU support in DirectML is currently only compatible with a subset of machine learning models that have been targeted for support. Some models may not run at all or may have high latency or low accuracy. We are working to improve the compatibility and performance of more models in the future.

The release encourages user feedback to enhance the effectiveness of NPU support in DirectML, emphasizing its status as a work in progress. The aim is to provide developers and users with a glimpse into the potential of this integration while acknowledging ongoing efforts to expand coverage for additional machine-learning models suitable for NPU acceleration in future updates.

More here.