Unused power headroom in the training and inference fleets studied by POLCA
~3% training / ~21% inferenceobserved
| Value kind | observed — Reported measurements, counts, and specifications keep the precision and scope stated by their source; an exact specification is not treated as a range. |
|---|---|
| Scope | Headroom is 100% minus the row-level peak power utilization POLCA reports: 97% for training, 79% for inference (Patel et al., ASPLOS 2024), in historical fleets. Low headroom does not mean small swings: in the same table, training's largest 2-second power spike was 37.5% of provisioned power against 9% for inference. POLCA separately simulated 30% additional provisioned servers under specific controls. Neither headroom figure is a deployable allowance for a 2026 reasoning, MoE, or disaggregated fleet: derive that from the proposed fleet's measured coincident demand, its tested protection and capping response, and the serving degradation you will allow. |
| As of | 2024-04 |
| Source | Microsoft Research / ACM, Patel et al., Characterizing Power Management Opportunities for LLMs in the Cloud, ASPLOS, April 2024; publisher page gives month precision. · Table 2 (peak power utilization; largest 2 s spike), fleet headroom characterization and §§6.4–6.6; 30% more servers is a separate simulation result |
| Review | checking…review by 2027-04-27 · standard cadence |
| Recorded changes | last 2026-09-16 · 5 revisions tracked |
| Claim id | power-oversubscription-headroom-training |
Where the guide uses it
- 13.6 Level 5 Integrated Systems Testing (IST) & Failure-Mode Demonstration
- 14.7 Capacity, Power & Thermal Management in Operation
- 16.3 Software, Orchestration & Efficiency at the Frontier
← Full numbers register — every date-stamped figure in the guide, with revision history.