DanS 20cbad687d Redesign benchmark to measure sustained (thermally throttled) hashrate ...

instead of initial burst performance. Previously the benchmark used a
fixed 20s warmup + 10s peak measurement, which reported inflated
results on thermally constrained hardware (e.g. 179 H/s vs actual
sustained 117 H/s on a MacBook Pro).

- Adaptive warmup with stability detection: mine for at least 90s,
  then compare rolling 10s hashrate windows. Require 3 consecutive
  windows within 5% before declaring thermal equilibrium (cap 300s)
- Average-based measurement: record mean hashrate over 30s instead
  of peak, reflecting real sustained throughput
- Start candidates at half the system cores — lower thread counts
  are rarely optimal and waste time warming up
- Add CoolingDown phase: 5s idle pause between tests so each starts
  from a similar thermal baseline
- Adaptive time estimates: use observed warmup durations from
  completed tests to predict remaining time
- UI shows Stabilizing when waiting for thermal equilibrium past
  the minimum warmup, Cooling during idle pauses"

2026-04-06 13:51:56 -05:00

143 KiB

Raw Blame History

View Raw