Parallax
Benchmarks

Measured on the device, frame by frame.

Every number here is frame-to-decision latency on shipping hardware, over a full day of real camera traffic. The method is below, and so is the harness, so you can check our work.

7.8 msMedian, reference graph on dock-03
9.6 msp99 over the same 24-hour run
128 fpsSustained, no frames dropped
23×Faster than a 180 ms cloud round trip
Frame budget

Where 8 ms sits against everything else.

A blink takes about 100 ms. A cloud round trip over a site link takes 180 to 420 ms. The whole reference graph finishes before the next frame at 128 fps.

Latency

Every model on every chip.

Median milliseconds, int8, batch 1. The reference graph is Depth-S, Depth-Track and a zone policy, the setup running on dock-03.

ModelApple Silicon
A18 Pro
Qualcomm Hexagon
QCS8550
NVIDIA Jetson
Orin NX 16 GB
Hailo
Hailo-8
Ambarella
CV72S
Reference graph5.97.84.86.68.9
Depth-S4.86.43.85.37.3
Depth-M11.214.68.112.916.3
Depth-Track0.91.20.81.11.4
Depth-Pose7.49.85.58.611.0
Depth-Stereo9.612.46.913.814.1
Depth-Event3.24.12.8n/an/a

Fastest chip per row in green. Depth-Event needs an event-camera interface, which Hailo-8 and CV72S boards do not expose yet.

Methodology

How we measure.

Benchmarks that hide batching or warm caches say little about a camera. These rules apply to every number we publish.

Metric
Frame-to-decision: from the frame leaving the sensor to the event being available to application code. Pre-processing and post-processing are included.
Load
One camera stream at its native frame rate, batch 1. No batching, no warm cache between runs, no frames skipped by the harness.
Duration
Each cell is the median of a 24-hour run. p99 is reported next to it on the same run, never from a separate, shorter one.
Thermals
Devices in their production enclosures at 25 °C ambient, passive cooling where the product ships with it.
Precision
Int8 with exposure-banded calibration. Accuracy figures on the models page are measured on the same int8 build.
Reproducibility

Run it yourself.

The harness ships with the SDK. Point it at a supported device and a camera, and it runs the same graph we report.

# install the CLI, then benchmark the reference graph on a connected device
$ pip install parallax-sdk
$ parallax bench --graph reference --device qcs8550 --camera dock-03 --duration 24h

  node          p50      p99
  depth-s       6.4 ms   7.9 ms
  depth-track   1.2 ms   1.6 ms
  zone-policy   0.2 ms   0.3 ms
  frame         7.8 ms   9.6 ms   128 fps · 0 dropped
  • Same harness we use internally, open source, one command per chip.
  • Traces export to Studio so you can see where each frame spent its budget.
  • Results are signed with the device ID and Runtime version, so runs are comparable.

See the numbers on your own camera.

Install the SDK, run parallax bench on your device and compare the trace with ours.