After much anticipation and delay, Microsoft is edging closer to the release of its Windows Copilot Runtime with the arrival of a developer preview for the Arm build of its DirectML AI tool. While still not fully production-ready, this release provides a valuable opportunity for developers to start experimenting with local AI applications, particularly for the Copilot+ PCs equipped with Qualcomm Hexagon neural processing units (NPUs). This move signals that the integration of AI into the edge computing space is rapidly approaching, although there are still some challenges to overcome.
Bringing AI capabilities to the edge requires specialized hardware accelerators, and the Copilot+ PCs have been promised with a robust AI accelerator capable of 45 TOPs (tera operations per second). This developer preview provides access to this accelerator alongside a suite of bundled AI models and tools designed to help developers begin building their own AI applications. Despite delays in the release of certain tools, such as the Win App SDK APIs for Microsoft’s Phi Silica model and the absence of PyTorch support, this release marks an important first step in making edge-based AI development a reality.
At the core of this DirectML Arm release is a new driver for the Copilot+ PCs’ NPU, which is designed to work seamlessly with an updated version of the Open Neural Network Exchange (ONNX) runtime. This integration supports a selection of open-source models from Hugging Face, with Microsoft also providing a repository of DirectML-ready models that have been specifically optimized by Qualcomm for use with this new runtime. While this sets the stage for future AI development, it’s important to note that this initial release comes with significant limitations.
Unfortunately, this first iteration is far from perfect. It suffers from several severe bugs that make it challenging, if not impossible, to use effectively in many cases. Additionally, the small selection of supported models limits the types of applications developers can currently build with the new tools. Despite these hurdles, the inclusion of DirectML in the Windows Copilot Runtime, along with its compatibility with the Arm-based Copilot+ PCs, represents a major milestone. DirectML is part of the DirectX family, which facilitates access to GPUs and NPUs, enabling developers to leverage hardware acceleration in their C++ and .NET applications. As the ecosystem matures, this foundational support will pave the way for more powerful AI-driven applications on edge devices.