Performance & Benchmarks

Empirical performance benchmarks measured across execution providers and browser configurations.

---

Model Latency Summary

Measured warm on a modern laptop CPU using onnxruntime-web's single-threaded WebAssembly execution provider (bun run bench):

Slot	Model	Input Dimensions	Weight Size	Warm Latency (WASM SIMD)	Warm Latency (WebGPU)
Face	YuNet 2023mar	[1, 3, 640, 640]	232 KB	51 ms	18 ms
Head Pose	MobileNetV3-Small	[1, 3, 224, 224]	5.8 MB	13 ms	6 ms
Gaze	MobileOne-S0 (L2CS)	[1, 3, 448, 448]	4.7 MB	104 ms	28 ms
Objects	YOLOX-Nano	[1, 3, 416, 416]	3.5 MB	75 ms	22 ms

---

Frame Budget Analysis

A naive pipeline running all four models on every frame in single-threaded WASM would require: $$\text{Total} = 51 + 13 + 104 + 75 = 243\text{ ms } (\approx 4.1\text{ FPS})$$

Vigilo WASM optimizes this through cadence decoupling:

1. Object Decoupling (objectHz: 1): Objects run only once per second, removing 75 ms from 9 out of every 10 frames.
2. Gaze Sub-sampling (gazeEvery: 2): Gaze runs only every 2nd face frame.
   • Odd Frames (Face + Pose): $51 + 13 \approx 64\text{ ms } (\approx 15.6\text{ FPS})$
   • Even Frames (Face + Pose + Gaze): $51 + 13 + 104 \approx 168\text{ ms } (\approx 5.9\text{ FPS})$
   • Average Frame Latency: $\approx 116\text{ ms } (\approx 8.6\text{ FPS})$

With WebGPU enabled, the average frame latency drops to ~25 ms, achieving a buttery-smooth ~30–40 FPS while keeping the CPU free for exam content.

---

Float32 vs Quantized int8 YuNet

Metric	YuNet (Float32)	YuNet (int8 Quantized)	Ratio
File Size	232 KB	110 KB	0.47x
WASM Latency	51 ms	209 ms	4.10x slower

WARNING

While quantized int8 models save ~120 KB of download bandwidth, they run 4x slower in onnxruntime-web's WASM backend due to unvectorized dequantization. Always deploy the float32 variant.

---

Running Benchmarks Locally

You can benchmark each model directly on your machine:

bun run bench

This runs cold-start and warm-start inference across all four models, reporting p50, p90, and p99 latencies.