gpuwait report

target http://localhost:11434, model qwen3:4b, generated 2026-09-30 08:12 UTC

request idle 38.5%
38.5% idle

Request Idle Score at concurrency 1. This is the percentage of the test window where the server had zero requests in flight (or, with a hardware sampler, where the GPU itself was not active). A high number here at concurrency 1 reflects one user's think-time between messages, not a broken server. See idle percent drop as concurrency rises below.

sampler: ollama-timing (no passwordless sudo for powermetrics, falling back to request-timing inference (run `sudo visudo` to add a NOPASSWD rule for powermetrics if you want hardware samples instead))

Idle percent by concurrency

concurrency 1
38.5%
concurrency 16
0.0%

Full results

concurrencyidle %busy %tok/s p50 latency sp95 latency sokfailed
138.561.525.63.0973.262120
160.0100.038.847.63253.38340