synchronized GPU power swing between the compute and communication phases of a large training job, at site or multi-site scale
tens to hundreds of MWestimate
| Value kind | estimate — Interpret this value according to its displayed kind, scope, source, and as-of date. |
|---|---|
| Scope | The authors' estimate of swing size at scale, not a measured ramp rate. Their concern is the swing's frequency, which can excite grid or generator modes. Meta reported swings of tens of megawatts on its Llama 3 training run (Chapter 4.8). |
| As of | 2025-08 |
| Source | NVIDIA/Microsoft/OpenAI, Power Stabilization for AI Training Datacenters (arXiv 2508.14318, Aug 2025) · arXiv 2508.14318 §I Introduction ("At scale, these swings can amount to tens or hundreds of megawatts"); §II-C (GPU power runs near TDP in compute and near idle in communication) |
| Review | checking…review by 2027-08-20 · standard cadence |
| Recorded changes | last 2026-10-16 · 3 revisions tracked |
| Claim id | synchronized-gpu-power-swing-per-load-step |
Where the guide uses it
← Full numbers register — every date-stamped figure in the guide, with revision history.