Protocol Lab
Measure the path from a control command to a visible response.
Join this benchmark lab
Use the invitation link from this Mac, or enter its temporary access key.
Checking capabilities…
Complete suite: 29 browser codec profiles, 20 native encode/decode profiles, then sequential video + control comparisons with Xavier, Spark, Thor and Spark / Quinn. Keep this tab visible. Results upload between trials, after their media connections stop.
The page reports readiness to the Mac. One browser runs at a time.
Checking the source and browser capabilities…
Network path
No connection measured yet.
The page host and the video route are separate. NAT can still allow a direct connection. Unavailable browser details are reported as unknown; a low or high RTT alone does not identify a relay.
Transport results · this session
| Transport / mapping | Round | Control p50 | Control p95 | Visible p50 | Visible p95 | Frames/s | Longest gap | Status |
|---|
Each trial saves its raw timing samples, selected-path evidence, configuration, missing responses, browser statistics, and errors. A lower latency with more loss or freezes is not automatically better.
Browser codec results
API submit → output callback; includes queueing and browser scheduling. Hardware preference is not proof of a hardware backend. Equal bitrate is not equal image quality.
| Codec / profile | Requested acceleration | Encode p50 / p95 | Decode p50 / p95 | Decoded / scheduled | During drain | Status |
|---|
Saved on the Mac
Review exact browser, source and profile before comparing. Swipe tables sideways on a phone. Different tests are not automatically matched trials.
What is compared
- WebRTC media, default and reduced receiver jitter targets.
- WebRTC data channel + WebCodecs for a common decoder comparison.
- QUIC: persistent reliable stream, one stream per frame, datagrams, deadline-limited datagrams, and separate control/media connections.
Measurement boundaries
Network RTT is measured entirely on the testing browser’s clock. Server processing time is measured entirely on the source computer’s clock. Estimated one-way frame age includes a reported clock-offset interval; it is never treated as exact.
The baseline compares five QUIC mappings on one engine. The Spark Quinn target adds an independent engine and ordered GOP streams. This does not establish a ranking of all QUIC implementations.