REALTIMEAI.VIDEORTR Methodology
Menu

RTR BENCHMARK · V1.0

A fast number is useless without its clock.

Our methodology keeps the start event, stop event, media duration, endpoint, resolution, queue, sample count, and evidence source attached to every result.

REAL-TIME RATIO
RTR=

generation timeplayable duration

RTR < 1.00 crosses the playback threshold

Two clocks, two answers.

Inference RTR

Starts when the model begins inference and stops when generation completes. It isolates model and serving throughput, but does not represent the whole user wait.

End-to-End RTR

Starts when the client sends the request and stops when a playable result is ready. It includes queueing, preprocessing, inference, encoding, storage, and delivery.

Class A protocol

Record request sent, inference start when exposed, inference end, first playable response, and actual decoded media duration. Report p50, p95, range, failures, credits, audio, seed behavior, and prompt expansion.

Class B protocol

RTR alone is insufficient. Record session setup, time to first frame, sustained output FPS or segment cadence, jitter, stalls, reconnection behavior, and whether control can change during generation.

Class C protocol

Record action-to-effect latency, accepted input rate, output FPS, state persistence, controllable actions, revisit consistency, session horizon, and the hardware or hosted tier.

Evidence levels

Independent test

RealtimeAI.video can reproduce the run and publishes conditions or raw output.

Official claim

A model creator or hosting provider reports the value in a first-party source.

Project repository

An open project documents its own hardware and performance results.

Corrections

Send a primary source, raw result, or correction to data@realtimeai.video. Material changes update the verification date and remain visible in version history.