AI 中文总结
研究当多独立人工智能训练作业共享功率包络时其周期能否同步的问题，通过将机群形式化为广义Kuramoto系统，确定负载相关节流为耦合通道，得出维度、检测、缓解三方面结论，双作业共上限测量可证伪预测。
AI 中文摘要
大规模人工智能训练使计算设施成为数兆瓦级负载，其功率消耗具有周期性：数以万计的加速器在接近峰值功率的计算密集阶段和空闲的通信密集阶段之间同步运行。先前工作将每个设施视为电网的外部周期性强迫。我们提出运营商的问题：当许多独立训练作业共享一个超额认购的功率包络时，它们的周期会保持独立，从而总波动随作业数量的平方根增长，还是功率管理堆栈能将它们锁相为线性增长？这是一群非线性振荡器中的新兴同步——Kuramoto 情形，但由于加速器时钟与线路频率解耦，不存在经典耦合。我们确定了负载相关节流中的耦合通道：当总需求高时，上限、电压降和共享冷却会精确减慢计算。将机群形式化为广义 Kuramoto 系统，我们得到三个面向运营商的结论。维度：耦合在主导阶是排斥的，仅当控制回路的相位滞后超过半个周期时才变为吸引；保护是模式选择性的，因此需要速率多样性。检测：频率相关的挫折使起始变为一阶且具有滞后性。缓解：相位散射调度会同时提高每个模式阈值。我们指定的双作业共上限测量可证伪该预测。
英文摘要
Large-scale AI training turns computing facilities into multi-megawatt loads whose power draw is periodic: tens of thousands of accelerators step in lockstep between compute-bound phases near peak power and communication-bound phases where they idle. Prior work treats each facility as an exogenous periodic forcing on the grid. We pose the operator's question instead: when many independent training jobs share one oversubscribed power envelope, do their cycles stay independent, so aggregate fluctuation grows as the square root of the number of jobs, or can the power-management stack phase-lock them into linear growth? This is emergent synchronization in a population of nonlinear oscillators - the Kuramoto setting - but classical coupling is absent, since accelerator clocks are decoupled from line frequency. We identify the coupling channel in load-dependent throttling: caps, voltage droop, and shared cooling slow computation exactly when aggregate demand is high. Formalizing the fleet as a generalized Kuramoto system, we obtain three operator-facing statements. Dimension: the coupling is repulsive to leading order and turns attractive only when the control loop's phase lag exceeds half a cycle; protection is mode-selective, so rate diversity is required. Detect: frequency-correlated frustration makes the onset first-order and hysteretic. Mitigate: phase-scattering scheduling raises every mode threshold at once. The prediction is falsifiable by a two-job co-capped measurement, which we specify.
Comments42 pages, 10 figures
