To run YOLOv8 on an Intel iGPU/NPU, the least-effort route is: export to OpenVINO format, then keep using the official ultralytics API. The NPU is abstracted as intel:NPU, and format conversion, pre-processing and post-processing are all handled. This guide goes from a raw yolov8n.pt all the way to hundreds of FPS.

Env: Core Ultra 5 235H / OpenVINO 2026.5.0.dev (devices: CPU/GPU/NPU) / yolov8n.pt → yolov8n_openvino_model (FP16). NPU latency ~31ms/frame. Tested 2026-09-12.

Why OpenVINO format

PyTorch's .pt is a training format — it infers on CPU but won't automatically touch the iGPU's Xe cores or the NPU. OpenVINO is Intel's intermediate representation (IR), compiling the model into an operator graph that runs on those two chips. It's less a 'model file' than a device-scheduling runtime ecosystem.

Two steps: export + predict

Step one exports yolov8n.pt to an OpenVINO folder (you get yolov8n_openvino_model/ holding FP16 IR files):

from ultralytics import YOLO

model = YOLO("yolov8n.pt")
model.export(format="openvino", half=True)  # FP16 IR 导出

Step two loads the exported folder and sets the NPU device. One flag matters: device="intel:NPU". For person detection alone use classes=[0] (COCO class 0); the default conf=0.25 behaved well on surveillance footage.

from ultralytics import YOLO

model = YOLO("yolov8n_openvino_model")
results = model.predict(
    source="frame.jpg",
    device="intel:NPU",
    conf=0.25,
    classes=[0],
)

The result object gives you classes, confidence and box coords directly — iterate results[0].boxes. That's the payoff of the official path: OpenVINO's letterbox, BGR ordering and NMS decode — fiddly, error-prone bits — are all handled inside ultralytics. You never hand-write any of it.

Feeding the NPU: one frame per second is enough

Real surveillance doesn't need per-frame inference. Pulling frames with ffmpeg instead of cv2 full-decode is the watershed for throughput:

:: 每秒抽1帧（n%%fps==1 采样），比 cv2 省 hwdownload 回读
ffmpeg -i "%%f" -vf "fps=1" -f image2pipe -c:v rawvideo -pix_fmt rgb24 pipe:1

:: 或者按 n%%fps==1 抽到内存管道：检测只用 1 帧/s

Quick math: a 60s clip yields 60 frames, and one NPU process takes ~13s (≈4.6 FPS incl. load overhead); with 6 parallel jobs, 60 files / 1h of video needs only ~150s — 24× realtime. For hundreds of × realtime, that's the hybrid GPU+NPU pool's lane — see the throughput ladder.

Two-way validation: no misses, no false alarms

Before trusting the model for real work, prove it doesn't fire randomly. Test both extremes — an empty 6am and a busy 7pm — covering false positives and misses in one pass: