# benchmarks

VAOS™ benchmarks

Runtime latency, throughput, and scaling data across reference platforms. All measurements at v0.9.x on default configuration unless noted.

Reproduce on GitHub →
# runtime latency

VAOS Runtime Latency

End-to-end scene contract latency from sensor frame to validated output. Measured at 30 fps per camera, full OSTVAL™ pipeline enabled.

Platform Cameras FPS Latency
Raspberry P5 + RK1820 2 30 45 ms
RK3588 + RK1828 8 30 30 ms
Jetson Orin 4 30 40 ms
RTX 3070 6 30 12 ms
# scene contract

Scene Contract Throughput

Scene generation throughput Projected · (v2.0 benchmark run pending) on RK3588 + Rk1828 · 4 cameras · 30 fps. Objects and events per second at sustained load.

Metric Value
Objects / sec 520
Events / sec 138
Memory usage 256 MB
# multi-camera fusion

Multi-camera Fusion

Fusion pipeline latency scaling across camera count. Reference platform: RK3588 + RK1828 · 30 fps per camera.

Cameras Latency
1 camera 12 ms
2 cameras 15 ms
4 cameras 21 ms
8 cameras 38 ms
# memory runtime

Memory Runtime

Storage impact of the VAOS Memory Engine over time. Reference platform: RK3588 + RK1828 · 4 cameras · default retention policy.

Duration Storage impact
1 hour 12 MB
24 hours 280 MB
7 days 1.9 GB
30 days 7.4 GB
# scenelm

SceneLM

Inference latency per scene contract on RK3588 + RK1828 NPU. Lower is better. Measured at single-scene, no batching.

Model Inference latency
Gemma 2B ~3s
Gemma 4B ~5s
Qwen 3 ~6s
Reproduce these benchmarks — all test scripts and hardware configurations are published in the vaos-online GitHub repository under benchmarks/. Community contributions welcome — see the contributors guide. (Some values are projected engineering targets based on current VAOS v2.0 runtime assumptions. Final measured results will be published with reproducible benchmark scripts and hardware configuration files)