Performance

Ratios travel. Absolutes are box-dependent.

Every number here is measured, and every number carries the box and the build it was measured on. Read the caption before you plan a deployment around a figure.

~5%

of a 20-core NAS for the whole stack, under 1 GiB of RAM

2.4%

of one core to ingest 16 RTSP streams, hub only

0.064 ms

p95 bus latency, authenticated TCP on loopback

Measurement	Result	Box and build
Whole-stack footprint	~1.0 / 20 cores (~5%), <1 GiB	QNAP TVS-h874, 8 cameras / 16 streams, production
Ingest cost, hub only	~2.4% of one core	Same box, 16 RTSP streams
Detection, GPU	2.05 ms / inference	yolo11s @ 640, RTX 5090, dev box
Detection, OpenVINO on CPU	8.95 ms / inference	yolo11s @ 640, dev box — ~2.9× faster than the default CPU provider
Detection, CPU	25.85 ms / inference	yolo11s @ 640, ONNX Runtime CPU provider, dev box
Bus latency, authenticated TCP	p95 0.064 ms	Loopback, debug build
Bus latency, over TLS	p95 0.210 ms	Loopback, debug build
Transport floor	UDS ~5× loopback TCP	Synthetic, 1 KiB frames

Why it is this small

One epoll thread in the core, raw syscalls, no libc, and each camera ingested exactly once regardless of how many consumers want it. Recording is motion-gated, so disk cost follows events rather than wall-clock time.

How to size a box

Ingest scales with stream count and is cheap. Detection dominates, and the provider ladder decides how much: a GPU, an Intel iGPU or plain CPU, in that order of appetite. Pick the rung, then the box.

Reproducing this

The benchmark board is part of the documentation set, with the method next to each figure. Latency figures come from a debug build on loopback, which is the honest way to read them: a floor, not a promise.