Closed-Form Energy-Optimal Rejuvenation Interval T* & CAETOM Operating-Point Selector
| Model | Operating Point | 48h Decode Drift | Warm Rejuv Overhead | Optimal T* (Hours) | Paper 4 Policy Decision |
|---|---|---|---|---|---|
| phi3:mini | R1 (Unthrottled, 4.1 GHz) | -15.5% | 839.5 J | 1.67 | Periodic Warm Runner Restart every 1.67h (or cap at 75% + Warm Restart every 3.55h) |
| phi3:mini | 75% Cap (2.6 GHz — Peak T_J = 0.245 tok/J) | -7.4% | 839.5 J | 3.55 | 75% CPU Cap + Warm Runner Restart every 3.55h |
| qwen2.5:0.5b | 75% Cap (2.6 GHz — Peak T_J = 1.232 tok/J) | -0.3% (No Aging) | 155.6 J | INFINITY (No Rejuvenation Needed) | Operate on Fixed Schedule at 75% Cap (Rejuvenation incurs pure overhead) |
| llama3.1:8b | 62.5% Cap (2.1 GHz — Peak T_J = 0.145 tok/J) | -0.2% (No Aging) | 340.9 J | INFINITY (No Rejuvenation Needed) | Operate on Fixed Schedule at 62.5% Cap (Rejuvenation incurs pure overhead) |
| deepseek-v2:lite | 75% Cap (2.6 GHz — Peak T_J = 0.324 tok/J) | -0.5% (No Aging) | 872.8 J | INFINITY (No Rejuvenation Needed) | Operate on Fixed Schedule at 75% Cap (Rejuvenation incurs pure overhead) |