Windows ML Overview

Windows ML Overview#

This section describes how to use Windows ML on AMD Ryzen AI PCs and complements the official Microsoft Windows AI documentation. It bridges Microsoft’s local AI platform with Ryzen AI hardware and software.

Microsoft provides a comprehensive AI platform for Windows spanning three pillars:

  • Windows AI APIs: Built-in system APIs (OCR, image description, super resolution, object erase, etc.) for Copilot+ PCs. Use when your scenario is covered by these APIs.

  • Foundry Local: On-device runtime for LLMs and generative AI; auto-detects hardware and downloads compatible models. Use for LLM scenarios with minimal setup.

  • Windows ML: Runtime for custom ONNX models with automatic execution provider (EP) management across CPU, GPU, and NPU. Use when you need to run your own models.

On Ryzen AI PCs, Windows ML can leverage the NPU via the VitisAI EP (Execution Provider).

When to Use Windows ML#

Choose Windows ML when you:

  • Need to run custom ONNX models (CNN, Transformer, or LLM) on Windows

  • Want automatic EP management Windows downloads and registers compatible execution providers (VitisAI EP, MIGraphX EP, DirectML EP) on demand

  • Prefer C#, C++, or Python with a shared Windows-wide ONNX Runtime (smaller app size)

  • Need hardware flexibility select CPU, GPU, or NPU via execution policy

Use Windows AI APIs when built-in capabilities (OCR, image description, etc.) cover your scenario. Use Foundry Local when you want LLMs with minimal model preparation. Use the Ryzen AI NPU-only flow (Model Compilation and Deployment) when you need full control over ONNX Runtime without the Windows ML stack.

External Resources#