Benchmarks
How we will measure
The definitions first, in public, before any result exists. When a measurement is run it is published here with the hardware it ran on, the date and the method beside it. Nothing appears on this page that has not been measured.
What each measurement means
- Cameras per box
- The number of simultaneous 1080p streams at the stated frame rate a box sustains for one hour with detection on every frame sampled, CPU or NPU under 85%, no dropped events.
- Watts per camera
- Wall-plug power of the box under that load, divided by the number of streams.
- False-alarm rate
- On a held-out set of real site footage, the share of alerts a reviewer marked as not the event described. We publish the set size and the reviewer count.
- Latency
- Time from the frame in which the event begins to the alert leaving the box, median and 95th percentile.
What we will measure on
Small fanless x86 box
The default for a site of twenty cameras or fewer.
ARM board with an NPU
The low-power option, and the one we expect to be hardest on latency.
Existing NVR with spare compute
Measured alongside recording, not instead of it.
Cloud instance
For comparison only. It is not how we expect anyone to run this.
How results will be published
Benchmarks are versioned. Once a row is measured it stays on the page, older versions and all. If a competitor publishes a comparable measurement we will link it.
Every result will name the model version, the footage it was measured on, the size of that set, and who reviewed it. A number without those four things is not a benchmark, and we would rather show you an empty page than one.