DirectML

DirectML is Microsoft's GPU-agnostic ML acceleration API, layered on DirectX 12. It works on any Windows-supported GPU — NVIDIA, AMD, Intel, Qualcomm — without requiring CUDA or ROCm installs.

ONNX Runtime and a handful of inference engines support DirectML as a backend. Performance is generally 60–80% of vendor-native paths but installation is dramatically simpler — no driver toolkit, no version-matching pip wheels.

For local AI on Windows with a non-NVIDIA GPU (especially an Intel Arc or recent Snapdragon X laptop), DirectML is often the path of least resistance.

A Windows user with a Snapdragon X Elite laptop wants to run local inference but has no NVIDIA GPU and no ROCm-supported AMD card — just an integrated NPU/GPU combo that vendor ML stacks barely acknowledge. Rather than fighting driver installs, they run an ONNX Runtime build with the DirectML execution provider, which talks to the GPU through DirectX 12 and works out of the box since DirectX is already present on any Windows machine. The model — a small ONNX-exported Whisper for transcription — runs noticeably slower than the same model would on a CUDA-equipped desktop, but it runs at all, with zero toolkit setup. For laptop-class Windows hardware outside the NVIDIA/AMD-discrete lane, DirectML is often the only backend that doesn't require a driver archaeology project.