Microsoft is intensifying its efforts in AI by extending DirectML support to Copilot+ PCs, which rely on Qualcomm Snapdragon X Elite Compute Platform. The move is aimed at bolstering machine learning performance on ARM64-based devices.
Boosting AI on Copilot+ PCs
Initially introduced with Windows 10 version 1903, DirectML is a powerful API that enables machine learning workloads on any DirectX 12 compatible GPU, such as those from AMD, Intel, and NVIDIA. The recent update now extends this support to Copilot+ PCs featuring Qualcomm Hexagon NPUs. Necessary system requirements for this include DirectML version 1.15.2 (ARM64), ONNX Runtime version 1.18 (ARM64), Windows 11 version 24H2 or newer, and Qualcomm Hexagon NPU Driver version 30.0.31.250 or later.
According to Upendra Kulkarni, VP of Compute Software Product Management at Qualcomm, the Snapdragon X Elite's NPU offers 45 TOPS of AI performance and is designed for power efficiency. DirectML's DirectX API framework makes it easier for developers to migrate their models from GPU to NPU seamlessly.
WebNN Adoption for Web Apps
Moreover, Microsoft revealed that web developers can now tap into NPU power using DirectML and WebNN through the latest Developer Preview. WebNN, a new standard for running ML models in browsers, will enable more efficient AI-based web applications. The update adds preliminary compatibility with Intel Core Ultra processors featuring Intel AI Boost and Copilot+ PCs using Qualcomm Hexagon NPUs.
To utilize WebNN on DirectML, developers must ensure their systems are running Windows 11 version 24H2 or newer, have an Insider version of Edge, and possess updated drivers from WebNN NPU partners, including the Intel Core Ultra NPU and Qualcomm Hexagon NPU drivers.
DirectML is integral to the Windows Copilot Runtime, providing developers the ability to scale AI across various Windows devices. Those interested in this functionality must register an account on the Qualcomm Package Manager site to download the latest Qualcomm Hexagon NPU driver package and follow specific installation steps.
Redistributable Package and Upcoming Python API
Developers integrating DirectML 1.15.2 as a redistributable package in C++ applications should note that NPU selection via the Python API for ONNX Runtime is not available yet but is expected soon. Though specific ONNX models are assured to work on Qualcomm Hexagon NPUs, errors may occur with unsupported models.
Testing models not listed as supported can be done by enabling Windows developer mode, keeping in mind they might show undefined behavior. The WebNN Developer Preview allows testing ONNX models in the browser with hardware acceleration, requiring minimal code adjustments.
