AI inference glossary
Benchmark metrics

P90 interactivity

Also known as P90 TPS, P90 token rate, P90 streaming speed

In plain English

P90 interactivity expresses a slower-tail streaming interval as a token rate, so higher values mean faster response streaming.

Technical definition

In the Rubin AgentX analysis, P90 interactivity is the reciprocal of P90 full-response inter-token latency, expressed in tokens per second per user.

Relationship

P90 interactivity = 1000 / P90 full-response ITL (ms)

Engineering details

Take the P90 latency statistic first, then invert it with the appropriate unit conversion. A P90 full-response inter-token latency of 10 milliseconds corresponds to 100 tok/s/user. This is not the 90th percentile of token rates: reciprocation reverses the ordering of positive values.

Why it matters

The percentile and latency definition determine what speed target a comparison enforces. This metric excludes the initial wait before streaming; equal P90 interactivity can coexist with very different time to first token and end-to-end latency.

How to read it in InferenceX

The Rubin article compares engine-specific frontiers at matched P90 interactivity and interpolates only within measured ranges. Preserve the target and the comparison engine when citing its throughput, cost, or power ratios.