P50 latency
Also known as P50 latency, AI P50 latency
Definition
P50 latency is the serving concept concerned with p50 latency during AI inference.
What p50 latency means
P50 latency is the serving concept concerned with p50 latency during AI inference. In a serving system, define the measurement boundary before collecting data. State whether the worker is warm, which model and precision are active, how batching is configured, and whether the request is streamed. These details determine whether two similar measurements describe the same operating condition.
How it appears in a workload
An interactive coding request can record p50 latency together with prompt size, output size, route, and timestamps. Record relevant timestamps and runtime signals with the request rather than relying on one dashboard aggregate. A useful comparison includes request mix, arrival pattern, concurrency, and output limits. For routed traffic, retain the selected route so a change can be investigated rather than attributed to p50 latency without evidence.
Limitations and tradeoffs
The result depends on prompt shape, output length, concurrency, batching, and hardware. Use it with p50 and tail latency, useful token throughput, errors, memory pressure, and cost. The best configuration is the one that meets the application's reliability and responsiveness requirements under representative traffic.
How this relates to Weave
Weave Router comparisons can record p50 latency beside model, provider, workload, and request outcome. That context helps explain whether a routing change altered p50 latency, but Weave does not make an inference runtime faster by itself. Keep serving conditions visible when using these observations to compare complete AI tasks.
Explore Router