Do Co-Located AI Training Jobs Synchronize? Load-Dependent Throttling as a Coupling Mechanism for Phase-Locking Behind a Shared Power Cap

What the math gives to ML

The paper provides a concrete synchronization mechanism for populations of oscillators coupled through load-dependent throttling rather than through a physical clock: shared caps make each oscillator's phase velocity depend on aggregate demand. Its transferable asset is the signed effective coupling obtained by convolving a periodic workload waveform with a feedback response; phase lag can change coupling from desynchronizing to synchronizing, while hard saturation creates a separate nonlinear failure mode. A useful neural-network translation is a phase-aware controller for distributed workers or parameter blocks whose update rate, learning rate, or microbatch budget is modulated by shared resource pressure. The key test is not merely final accuracy: the controller should shift the measurable synchronization boundary at the predicted phase-lag condition and reduce the order parameter of worker update phases when operated in the repulsive regime.