Latency
End-to-end delay from a frame arriving at the input to being available at the output.
Docs
Why FPS Is Not Enough
A methodology for evaluating real-time video processing pipelines.
The thesis
FPS tells only part of the story
Average frames-per-second is easy to optimize and easy to flatter. When multiple concurrent streams share decode and inference resources, what matters is whether frames are processed in time, whether the pipeline recovers after stalls, and whether latency stays stable. This methodology documents how we evaluate real-time video pipelines systematically.
Test Platform
What we ran it on
Edge evaluation runs on an RK3588 platform, matching close-to-deployment conditions.
Input
Test input source
We use streams encoded by real cameras rather than synthetic data, so decoder pressure reflects reality.
Concurrency
Concurrent streams
Scaling from one stream to many, observing how behavior changes under shared resources.
Metrics
What we measure
Each metric maps to a real risk: whether processing is real-time, whether latency is stable, whether the pipeline recovers, and whether the decoder holds up.
Frame Age
How far the currently processed frame lags behind the live source.
P99
The 99th-percentile latency — whether tail latency stays stable.
Burst Tolerance
Behavior under traffic bursts or bandwidth fluctuations.
Recovery
How quickly the pipeline recovers after a stream break or decoder failure.
Decoder Stall
Decoder stalls, dropped frames, or resource exhaustion.
Results
Results & data
Placeholder — to be filled with measured curves and data tables for each concurrency level.
Get Started
See the platform in action
Explore the technical demo, or tell us about your use case to request a live walkthrough.