Overview
A playground of small, pokeable utilities for the laptop’s built-in Intel AI Boost NPU (Core Ultra 9 275HX, ~13 TOPS). The point isn’t raw speed -; the discrete GPU out-throughputs it every time -; it’s power and thermals: the NPU can sip ~2 watts doing always-on vision while leaving the GPU free. It also links to the Reolink cameras around the house for realtime analysis.
How It Works
Everything uses the OpenVINO workflow that mirrors the Hailo edge-AI work: a model is compiled ahead of time, then runs cheap and cool on the accelerator. A --device flag flips the same pipeline across NPU, the integrated GPU and CPU so you can watch FPS and package power change. The toolbox:
01_hello_npu device caps + NPU vs iGPU vs CPU benchmark 02_detect real-time object detection (YOLO11 → OpenVINO) 03_pose skeleton overlay (YOLO11-pose) 04_face face boxes + 5-pt landmarks (YuNet ONNX) 05_audio live mic → text (OpenVINO Whisper) audio_bark sound-event detection (YAMNet, split CPU+NPU)
Reolink Integration
Every vision script takes a --source, so handing it an RTSP URL points it at a house camera instead of the webcam. The low-res _sub substream is ideal for continuous NPU inference -; less decode and bandwidth, still plenty for detection -; and --classes 0 16 keeps only people and dogs. The natural next build is an “is anyone (or any dog) in the driveway” watcher that fires a snapshot when a box persists, wiring the hit straight into the existing Telegram notifier.
Current Status
Five examples plus the audio sound-event detector are working. It’s a learning bench more than a product -; the goal is fluency with the NPU toolchain and a reusable npu_common.py for the always-on camera ideas.