Energy per token, for local AI.
Lemonade servers already report tokens/sec and latency. Lemon Metrics adds the missing layer: how many joules each token costs on your CPU, iGPU, or NPU — so the battery fight between an NPU and an iGPU on the same die becomes explicit.
How it works
Measure energy
A power-sampling layer wraps lemonade bench, recording wall-clock
watts at 1–2 Hz from hwmon, RAPL, rocm-smi, or powermetrics.
Compare hardware
One leaderboard per device × model × backend. Sort by any metric and compare CPU, iGPU, and NPU runs side by side.
Vote for truth
Every result is community-verified. Up-vote runs you trust, down-vote anything that looks wrong — flagged runs are hidden until re-verified.
Leaderboard
One row per device × model × backend. Best = lowest J/token; perf-only runs fill the other columns. Sort by any column; each row links to its full report. Explore and compare all results →
| Device | Model | Backend | J/token | tok/s | peak W | tok/kWh | Votes | when |
|---|
Why joules?
With rising costs of energy and laptops with battery duration, the real competition is between the CPU, integrated GPU, and NPU sitting on the same die. An NPU that runs 30% slower but draws half the power which means it wins the battery fight — but “best tok/s” alone hides it in pursuit for speed. J/token makes that tradeoff explicit, and with Lemon Metric the raw power samples are published with every run.