Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Co-located AI training jobs behind a shared power cap can synchronize—or be actively spread apart—depending on one number: the controller's lag at the iteration frequency.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 12:11 UTC pith:AWVSPVCI

load-bearing objection A clean derivation of the coupling sign, a solid numerical engine—and an abstract that overstates the first-order onset. the 3 major comments →

arxiv 2607.19638 v1 pith:AWVSPVCI submitted 2026-07-22 eess.SY cs.DCcs.SYnlin.AO

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

Brieuc Le Roux Tardif This is my paper
classification eess.SY cs.DCcs.SYnlin.AO MSC 34C1534D06
keywords power cappingload-dependent throttlingKuramoto modelSakaguchi frustrationphase synchronizationAI training workloadscontrol delayhysteresis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper asks whether many AI training jobs sharing one oversubscribed power envelope keep their power cycles independent—aggregate fluctuation growing as the square root of the number of jobs—or phase-lock through the site's own power management, making the site swing coherently with amplitude growing as the number of jobs. It claims the coupling channel is load-dependent throttling: caps, voltage droop, and shared cooling slow computation exactly when aggregate demand is high. Reducing the fleet to a generalized Kuramoto oscillator model, it derives a sign law in which the coupling is repulsive when the control loop's phase lag is below half an iteration period and attractive above it, with the critical delay being one half of an iteration. It further claims the onset is first-order and hysteretic for diverse fleets, and that homogeneous fleets can lock at higher harmonics even when the fundamental mode is repulsive. If the paper is right, operators can avoid coherent grid-visible swings by keeping capping fast and by deliberately diversifying iteration rates.

Core claim

The central claim is that the aggregate power demand of many co-located training jobs does not stay statistically independent: the power-management stack itself supplies a coupling channel. Because throttling acts only while a job is compute-bound, it acts as a phase-dependent force rather than a common slowdown, closing the jobs into a self-consistent dynamical system. For a narrow spread of iteration rates and weak throttling, this reduces to a Sakaguchi–Daido (frustrated Kuramoto) coupling with frustration α = ω̄τ + π/2, so the effective synchronizing coefficient is a1 = −sin(ω̄τ). Repulsion is the default; attraction—fleet-wide phase locking at the iteration frequency—appears only when t

What carries the argument

The central object is the load-dependent throttle gated by each job's compute phase: the susceptibility χ_i, active only during a job's compute-bound interval, converts a shared aggregate signal into a genuine per-phase force. Linearizing the throttle and averaging over an iteration yields a Sakaguchi–Daido coupling (a Kuramoto model with a constant phase lag), whose frustration α = ω̄τ + π/2 is set entirely by the control delay, and whose sign a1 = −sin(ω̄τ) decides whether phases repel or lock. The per-mode variant a(m) ∝ −sin(m ω̄τ) governs higher-harmonic cluster states.

Load-bearing premise

The load-bearing step is the slow-phase approximation that replaces a job's true delayed phase by its natural-rate phase lag, φj(t−τ) ≈ φj(t) − ωjτ; the paper's own tests show this substitution is imperfect (the averaged model predicts a delay-inflated threshold 2γe^{γτ}, while the full delay-differential model lands at 2γ), so if the substitution is not faithful, the frequency–frustration correlation behind the first-order onset is partly a reduction artifact.

What would settle it

Run two jobs under one capping controller with no other load and with the control delay below half an iteration, then measure the cross-correlation of their power traces; the theory predicts anti-correlation (repulsion), with the coupling sign following −sin(ω̄τ) and the phase offset Δφ between the power peak and cap susceptibility readable from the same measurement. If the traces are positively correlated at short delay, the mechanism's sign law is wrong. A complementary check: sweep the delay past half an iteration and observe the correlation flip to positive, which would confirm the predict

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Fast electrical capping (millisecond lag) sits deep in the repulsive window: the model predicts it actively spreads the fleet and holds aggregate fluctuation below the independent-phases baseline, not merely unamplified.
  • Attraction, and the resulting N-scaling coherent swing, requires accumulated dead time—polling intervals, sample-and-hold, cascaded slow stages—crossing half a period; a single analog thermal lag cannot invert the sign on its own.
  • No choice of control delay protects every mode: at τ ≈ T/4 the fundamental is maximally repulsive while the third harmonic is maximally attractive, so a sufficiently uniform fleet locks at three times the iteration frequency with the fundamental order parameter blind to it.
  • Because the onset is first-order and hysteretic for diverse fleets, a fleet below the forward threshold can be tipped into the locked state and stay there; the paper concludes operators should watch the order parameter and its harmonics in existing telemetry.
  • Deliberately detuning and phase-offsetting jobs (phase-scattering scheduling) raises every mode threshold at once, with a first-order throughput cost roughly equal to the fractional detuning spread.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A direct consequence the paper leaves implicit: if the sign law holds at rack scale, co-location density itself is a design variable—placing identical replicated jobs in one power domain is the most dangerous configuration, while mixing job types and iteration periods becomes a cheap architectural hedge.
  • The paper models only the speed (phase) channel of throttling; an inferred extension is that the simultaneous power-lowering amplitude channel, which the paper treats heuristically as damping, could itself carry lag and shift the effective frustration, so the measured sign in a two-job experiment may differ from −sin(ω̄τ) by a channel-dependent offset.
  • If the predicted repulsion is confirmed for fast loops, the framework suggests a testable mitigation: schedule jobs with deliberately spread iteration rates and measure whether the order parameter stays at the incoherent floor under a transient cap event; the paper's hysteresis result predicts a sticky locked branch for uniform fleets even after the disturbance passes.
  • The per-harmonic sign law suggests a spectral design rule beyond the paper's explicit scope: an operator could place the control delay so that the harmonics where the fleet is most uniform are repulsive, accepting attraction only at harmonics whose high thresholds make locking unlikely—the paper notes, however, that no single delay is repulsive at every mode.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper asks whether co-located AI training jobs sharing an oversubscribed power envelope can synchronize through load-dependent throttling, and formalizes the answer as a generalized Kuramoto system. Each job is a two-level power waveform with an internal phase; throttling is gated by the job's compute phase, and after linearization and rotating-wave averaging the dynamics reduce to a Sakaguchi–Daido coupling with frustration α = ω̄τ + π/2 and synchronizing coefficient a1(τ) = −sin(ω̄τ). Proposition 1 gives a sign reversal from repulsive to attractive at τ* = T/2; Proposition 2 gives a per-harmonic sign law and mode-m lock thresholds K_c^(m) = 2γ m²/|sin(mω̄τ)|; Proposition 3 gives the fundamental critical coupling K_c = 2/[πg(0)a1]. The paper further claims that the frequency-correlated frustration α_j = ω_j τ + π/2 makes the onset first-order and hysteretic for diverse fleets, and proposes phase-scattering scheduling as a mitigation. The numerical study integrates the unaveraged delay model and reports quantitative confirmation of the sign reversal, the leading-order threshold, the harmonic lock, the r√N amplification, and an opening hysteresis loop with rate spread.

Significance. If the results hold, the paper makes a substantial conceptual contribution: it identifies an operator-owned coupling channel—load-dependent throttling behind a shared cap—that can turn an incoherent aggregate (fluctuation ∝ √N) into a coherent one (∝ N), and it converts the mechanism into falsifiable predictions about loop delay, rate diversity, and harmonic locking. The derivation of the coupling coefficient and sign law is a genuine strength: it is derived rather than fitted, and the numerical engine is validated against the textbook Kuramoto onset Kc = 2γ before model-specific runs. The harmonic-lock threshold K_c^(3) = 18γ is derived in closed form, predicted before the sweep, and bracketed by the full-DDE simulation. The reproducibility apparatus (single orchestrator, fixed seeds, regenerable tables and figures) is exemplary and materially raises confidence in the reported numbers. The main gap is the load-bearing claim of first-order, hysteretic onset, which the paper itself leaves partly open and which is sensitive to the tail of the rate distribution.

major comments (3)
  1. [§3.4, Appendix B, Eq. (28), §6.2/§6.8] The slow-phase substitution φ_j(t−τ) ≈ φ_j(t) − ω_jτ (Assumption 5(i), Appendix B Step 2) is the source of the frequency-correlated frustration α_j = ω_jτ + π/2, which drives the abstract's 'Detect' claim of first-order, hysteretic onset. The paper's own full-DDE fine sweep, however, places the onset at the leading-order value 2γ, not at the averaged model's 2γ e^{γτ}; as §6.2 and Appendix D note, this means a locked cluster retards at a common frequency Ωτ, not at each job's natural rate. That is direct evidence that the frequency–frustration correlation may be partly a reduction artifact. Since the first-order transition in §4.3 is built on that correlation, the authors need either to derive a common-frequency correction to the frustration or to demonstrate subcriticality in the full DDE with a bounded-support rate distribution, before the operator-facing 'Detect' statement can stand.
  2. [§6.4, Table 3, §6.8] The hysteresis evidence is explicitly tail-sensitive. On a positive-truncated Lorentzian the γ = 0.2 loop closes to near the relaxation floor (gap 0.058 versus 0.68 on the unbounded Lorentzian), and the paper concedes that the first-order classification is 'protocol-bound and tail-sensitive' and that the wide loops are 'carried in part by the far tail of the unbounded Lorentzian.' Real iteration-rate distributions have bounded support and no negative rates, so the abstract's assertion that frequency-correlated frustration 'makes the onset first-order and hysteretic' is not established for realistic fleets. The revision should either supply bounded-support DDE evidence for subcriticality or explicitly restrict the Detect claim to heavy-tailed, unbounded rate families in the abstract, §4.3, and §7.
  3. [Appendix D, Eq. (30), §4.3] The Landau coefficient that would decide the order of the transition is left explicitly unevaluated, and the paper's evidence for the first-order onset is a finite-N warm-started continuation. This is an honest limitation, but it is load-bearing for one of the three headline operator-facing statements. As written, the analytical framework supplies only the normal-form structure and a conjecture of the form Reb = b*(1 − c(γτ)^2 + ...). The manuscript should either carry out the cubic closure on the OA manifold, or promote the first-order characterization to a stated numerical observation with the necessary caveats in the abstract, rather than presenting it as an established Detect statement.
minor comments (5)
  1. [§4.1, Eq. (16)] The coherence amplification factor r√N is introduced clearly, but the accompanying schematic (Figure 4) would benefit from an explicit statement that the locked-state amplitude is computed at the fundamental mode and that the same bookkeeping applies mode-by-mode; the text does say this, and a one-line addition to the caption would remove ambiguity.
  2. [§6.2, Table 3] The column heading 'onset' with entries 'continuous'/'first-order' is terse. Since the quantitative criterion is max|r↑−r↓|, adding that definition to the table caption would help readers interpret the protocol-bound classification.
  3. [§3.4, Eq. (12)] The statement that the reference phase φ⋆ cancels assumes exact alignment of the power and susceptibility fundamentals. The paper notes the relaxation at (12), but it would be helpful to state explicitly that the τ* = T/2 figure assumes Δφ = 0 in the one-sentence summary as well, not only in the derivation.
  4. [§6.8] The 'positive-rate fleet' check reports that the sign-law window means are unchanged and the onset marker is unchanged, but the hysteresis gaps are reshaped. Since this is a central caveat, consider making it a subsection of §6.4 rather than a bullet in the robustness section, so the reader of the main hysteresis discussion immediately sees the tail dependence.
  5. [§6.9] The reproducibility description is excellent. A small suggestion: list the Python dependencies (e.g., NumPy version) and the exact command to regenerate all outputs, so the 'single orchestrator' claim is independently verifiable without inspecting the source.

Circularity Check

0 steps flagged

No significant circularity: the coupling sign, thresholds, and harmonic-lock predictions are derived from stated assumptions and independently tested against the unaveraged DDE model.

full rationale

Walked the derivation chain from (9) to (11)-(14). The Sakaguchi frustration α=ω̄τ+π/2 and synchronizing coefficient a1(τ)=-sin(ω̄τ) are obtained by linearizing the throttle, substituting φj(t−τ)≈φj(t)−ωjτ, and retaining the slow term of the product-to-sum identity (Appendix B, Eq. 22); the predicted sign and critical delay are not put in by hand and are not defined in terms of phase-locking. The critical coupling Kc=2/[πa1(τ)g(0)] follows from the self-consistency integral (25) and is tested against the full DDE model, including a fine sweep that separates the leading-order 2γ onset from the averaged model's 2γe^{γτ}. Proposition 2's harmonic threshold Kc^(m)=2γm²/|sin(mω̄τ)| is derived per-mode and then used to make a quantitative prediction (γ*=K/18) that Table 5 brackets; this is a parameter derived before the sweep, not fitted to its outcome. Robustness checks include a textbook Kuramoto validation (r=sqrt(1-Kc/K)=0.71), plus time-step, seed, Fourier-truncation, and positive-rate-support tests. The explicitly open items—the Landau coefficient (30) and the bounded-support hysteresis question—are stated as limitations and do not function as circular support. No load-bearing self-citation occurs: reference [23] is an external antecedent for shear diversity, and no uniqueness theorem is imported from the authors. Thus no step reduces by definition, by fitted-input renaming, or by self-citation to its own conclusion.

Axiom & Free-Parameter Ledger

6 free parameters · 7 axioms · 0 invented entities

The paper is transparent about its assumptions (all numbered, with costs). No invented entities are postulated; the phase gate χi and the twisted order parameter are modeling constructs, not new forces, particles, or degrees of freedom. The main ledger items are: five domain assumptions that define the modeled system; the ad-hoc Lorentzian rate family used for exact solvability; and the folded coupling parameters η, h′, c1, χ1 which enter only through the swept knob K and are unmeasured — so the critical-coupling and phase-diagram predictions are dimensionless until calibrated. The first-order-onset claim additionally depends on the (averaged-model) frequency–frustration correlation whose faithful reproduction by the full delay system the paper's own fine sweep calls into question.

free parameters (6)
  • K (common coupling scale) = Swept 0.3–4.8 (normalized units) across experiments; physical value K = ½Nω̄ηh′χ1c1 never evaluated
    The central swept knob of the numerics. The four physical ingredients (throttle sensitivity η, cap stiffness h′, waveform contrast c1, gate fundamental χ1) are never measured and enter only through K; thresholds are dimensionless until the two-job experiment calibrates them (§7).
  • γ (Lorentzian rate half-width) = 0.002–0.5 per experiment
    Lorentzian half-width of the rate distribution, chosen for exact OA solvability; its unphysical negative-rate tail (14.8% of mass at γ=0.5) carries part of the measured hysteresis at intermediate spreads (§6.8).
  • δ (compute duty cycle) = 1/2 in main runs (0.3 in one earlier scan)
    Set to 1/2 because even harmonics vanish, making the harmonic-lock test sharpest; the paper notes that experiment establishes existence, not typicality (§6.6).
  • M (Fourier truncation) = 5 (1, 5, 15 in checks)
    Fourier truncation defining the integrated smooth model; Gibbs bounds ±0.09 at M=5, observables stable to sub-percent at M=15 (§6.8).
  • N (fleet size) = 300–400 main runs; 800 (E3, E8); 1200 (fine onset)
    Fleet size; order-parameter floor scales as N^{-1/2}; fine sweep at N=1200 resolves Kc to 0.100–0.101 (§6.2, §6.8).
  • τ (control delay) = Swept 0–2T in the delay studies
    The independent variable of the sign law; ω̄=1, T=2π normalization.
axioms (7)
  • domain assumption Assumption 1 — two-level power waveform; susceptibility gate χi shares the power waveform's duty cycle and phase
    §3.1. Real workloads overlap collectives with computation [19], raising P_lo and misaligning gate vs. power peaks; the paper notes this shifts the critical delay by Δφ/ω̄ and reduces K, and that multi-pulse waveforms break the m≥2 envelope.
  • domain assumption Assumption 2 — all-to-all mean-field throttling coupling
    §3.1. Real campuses are hierarchies of caps (block-local coupling); the paper states mean-field is a 'tractable worst-case candidate, not a proven bound' and flags modular-topology local resonances as open (Section 7).
  • domain assumption Assumption 3 — accelerator clocks decoupled from grid frequency
    §3.1, Fig. 1. Defines the problem: frequency cannot couple jobs; voltage acts through the throttling channel.
  • domain assumption Assumption 4 — throttle h monotone C², cap binds O(1) of the time (h′(s0) > 0)
    §3.1. If the cap binds only at peaks, the interaction is rectified and much weaker; the paper calls the rarely-binding case 'weaker (safer)'.
  • domain assumption Assumption 5 — weak coupling ε = K/ω̄ ≪ 1 and slow phases; cap saturation keeps K intensive
    §3.1, App. B. The reduction's error control; numerics track the full model to ε ≈ 0.6 and fail by ε ≈ 1.2. The 1/N cap-saturation scaling is 'a property of the controller law, not a theorem'. The natural-rate slow-phase substitution generates the frequency-correlated frustration.
  • ad hoc to paper Lorentzian rate density everywhere in the numerics
    Chosen so that OA closed forms are exact; its negative-rate tail (14.8% of mass at γ = 0.5) carries part of the measured hysteresis loops, and the paper notes truncated-Lorentzian loops close at intermediate spreads (§6.8).
  • standard math Standard Kuramoto self-consistency, linear stability of incoherence, OA manifold reduction
    App. C–D: the critical coupling (17), the dispersion relation (26), and the Stuart–Landau structure (29) rely on classical results [17,29,30,39].

pith-pipeline@v1.3.0-alltime-deepseek · 36978 in / 32699 out tokens · 324777 ms · 2026-08-01T12:11:21.876286+00:00 · methodology

0 comments
read the original abstract

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.

Figures

Figures reproduced from arXiv: 2607.19638 by Brieuc Le Roux Tardif.

Figure 1
Figure 1. Figure 1: The coupling mechanism, schematically (formalized in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The building blocks of the model, at the duty cycle [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The per-harmonic sign law a (m) (τ ) = − sin(mωτ ¯ ) for m = 1, 2, 3 (Propositions 1 and 2); positive values are attractive (locking), negative repulsive (splaying). The fundamental is repulsive across the whole window 0 < ωτ < π ¯ and reverses at τ ⋆ = T /2 (right dotted line), but no delay is repulsive at every mode: at ωτ ¯ = π/2 (left dotted line) the fundamental is maximally repulsive (filled dot) whi… view at source ↗
Figure 4
Figure 4. Figure 4: The stake of §4.1, schematically: N = 12 two-level jobs (δ = 1 2 ), phase configuration (top, with the order-parameter vector where it is macroscopic) and the exact aggregate deviation of that configuration (bottom, common vertical scale). Independent phases leave an incoherent ripple of scale √ N (r = O(N −1/2 )); the splay ordering favoured by repulsive coupling cancels the aggregate below the independen… view at source ↗
Figure 5
Figure 5. Figure 5: Order of the onset – an interpretive sketch, not a computed bifurcation diagram (the [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: E1 – measured order parameter r1 (dots, unaveraged DDE model) against the derived coupling coefficient a1(τ ) = − sin ¯ωτ (dashed) across one period of the control delay. The fleet is incoherent across the repulsive window and locked throughout the attractive window (shaded), with the reversal at the predicted τ ⋆ = T /2. Two elevated features are explained in the text: the exactly-reactive points (τ = 0, … view at source ↗
Figure 7
Figure 7. Figure 7: E3 – aggregate power fluctuation divided by [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: E7 – warm-started up/down continuation of the order parameter at [PITH_FULL_IMAGE:figures/full_fig_p025_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: E8 – the crossover of Table [PITH_FULL_IMAGE:figures/full_fig_p027_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: E9 – steady-state phase diagram over the control-delay/coupling plane, from the full DDE [PITH_FULL_IMAGE:figures/full_fig_p028_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Anonymous sharing is pairwise phase-blind

    eess.SY 2026-07 accept novelty 8.0

    Anonymous shared resources give identical checkpointing jobs no pairwise phase coupling; the two-job gap map is the identity, firing order freezes, and synchrony is unreachable and non-attracting.

Reference graph

Works this paper leans on

44 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Acebrón, L

    Juan A. Acebrón, L. L. Bonilla, Conrad J. Pérez Vicente, Félix Ritort, and Renato Spigler. The Kuramoto model: A simple paradigm for synchronization phenomena.Reviews of Modern Physics, 77(1):137–185, 2005

    2005

  2. [2]

    Synchronization in complex networks.Physics Reports, 469(3):93–153, 2008

    Alex Arenas, Albert Díaz-Guilera, Jürgen Kurths, Yamir Moreno, and Changsong Zhou. Synchronization in complex networks.Physics Reports, 469(3):93–153, 2008

    2008

  3. [3]

    Bergen and David J

    Arthur R. Bergen and David J. Hill. A structure preserving model for power system stability analysis.IEEE Transactions on Power Apparatus and Systems, PAS-100(1):25–35, 1981

    1981

  4. [4]

    Modal analysis of spatial load correlation in AI data center-dominated power systems

    Chandan Chaudhary, Michael Murillo, Mohammed Ben-Idris, Joydeep Mitra, Dilip Pandit, and Atri Bera. Modal analysis of spatial load correlation in AI data center-dominated power systems. arXiv:2606.13847, 2026

    Pith/arXiv arXiv 2026

  5. [5]

    Power stabilization for AI training datacenters

    Esha Choukse, Brijesh Warrier, Scot Heath, et al. Power stabilization for AI training datacenters. arXiv:2508.14318, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Hiroaki Daido. Onset of cooperative entrainment in limit-cycle oscillators with uniform all-to-all interactions: bifurcation of the order function.Physica D: Nonlinear Phenomena, 91(1–2):24–66, 1996. 39

    1996

  7. [7]

    Hanebutte, Rahul Khanna, and Christian Le

    Howard David, Eugene Gorbatov, Ulf R. Hanebutte, Rahul Khanna, and Christian Le. RAPL: Memory power estimation and capping. InProceedings of the 16th ACM/IEEE International Symposium on Low Power Electronics and Design (ISLPED), pages 189–194, 2010

    2010

  8. [8]

    Synchronization in complex networks of phase oscillators: A survey.Automatica, 50(6):1539–1564, 2014

    Florian Dörfler and Francesco Bullo. Synchronization in complex networks of phase oscillators: A survey.Automatica, 50(6):1539–1564, 2014

    2014

  9. [9]

    Synchronization in complex oscillator networks and smart grids.Proceedings of the National Academy of Sciences, 110(6):2005–2010, 2013

    Florian Dörfler, Michael Chertkov, and Francesco Bullo. Synchronization in complex oscillator networks and smart grids.Proceedings of the National Academy of Sciences, 110(6):2005–2010, 2013

    2005

  10. [10]

    Power provisioning for a warehouse- sized computer

    Xiaobo Fan, Wolf-Dietrich Weber, and Luiz André Barroso. Power provisioning for a warehouse- sized computer. InProceedings of the 34th Annual International Symposium on Computer Architecture (ISCA), pages 13–23, 2007

    2007

  11. [11]

    Filatrella, A

    G. Filatrella, A. H. Nielsen, and N. F. Pedersen. Analysis of a power grid using a Kuramoto-like model.The European Physical Journal B, 61(4):485–491, 2008

    2008

  12. [12]

    Explosive synchroniza- tion transitions in scale-free networks.Physical Review Letters, 106(12):128701, 2011

    Jesús Gómez-Gardeñes, Sergio Gómez, Alex Arenas, and Yamir Moreno. Explosive synchroniza- tion transitions in scale-free networks.Physical Review Letters, 106(12):128701, 2011

    2011

  13. [13]

    Clustering and slow switching in globally coupled phase oscillators.Physical Review E, 48(5):3470–3477, 1993

    David Hansel, Germán Mato, and Claude Meunier. Clustering and slow switching in globally coupled phase oscillators.Physical Review E, 48(5):3470–3477, 1993

    1993

  14. [14]

    MegaScale: Scaling large language model training to more than 10,000 GPUs

    Ziheng Jiang, Haibin Lin, Yinmin Zhong, et al. MegaScale: Scaling large language model training to more than 10,000 GPUs. In21st USENIX Symposium on Networked Systems Design and Implementation (NSDI), 2024

    2024

  15. [15]

    Wide-area power system oscillations from large-scale AI workloads

    Min-Seung Ko and Hao Zhu. Wide-area power system oscillations from large-scale AI workloads. arXiv:2508.16457, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Mitigation of datacenter demand ramping and fluctuation using hybrid ESS and supercapacitor

    Min-Seung Ko, Jae Woong Shim, and Hao Zhu. Mitigation of datacenter demand ramping and fluctuation using hybrid ESS and supercapacitor. arXiv:2512.08076, 2025

    arXiv 2025

  17. [17]

    Self-entrainment of a population of coupled non-linear oscillators

    Yoshiki Kuramoto. Self-entrainment of a population of coupled non-linear oscillators. In International Symposium on Mathematical Problems in Theoretical Physics, volume 39 of Lecture Notes in Physics, pages 420–422. Springer, 1975

    1975

  18. [18]

    Springer, 1984

    Yoshiki Kuramoto.Chemical Oscillations, Waves, and Turbulence, volume 19 ofSpringer Series in Synergetics. Springer, 1984

    1984

  19. [19]

    Charac- terizing compute-communication overlap in GPU-accelerated distributed deep learning: Perfor- mance and power implications

    Seonho Lee, Jihwan Oh, Junkyum Kim, Seokjin Go, Jongse Park, and Divya Mahajan. Charac- terizing compute-communication overlap in GPU-accelerated distributed deep learning: Perfor- mance and power implications. arXiv:2507.03114, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    AI load dynamics—a power electronics perspective

    Yuzhuo Li and Yunwei Li. AI load dynamics—a power electronics perspective. arXiv:2502.01647, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Recalibrating global data center energy-use estimates.Science, 367(6481):984–986, 2020

    Eric Masanet, Arman Shehabi, Nuoa Lei, Sarah Smith, and Jonathan Koomey. Recalibrating global data center energy-use estimates.Science, 367(6481):984–986, 2020

    2020

  22. [22]

    Mirollo and Steven H

    Renato E. Mirollo and Steven H. Strogatz. Synchronization of pulse-coupled biological oscillators. SIAM Journal on Applied Mathematics, 50(6):1645–1662, 1990. 40

    1990

  23. [23]

    Shear diversity prevents collective synchronization.Physical Review Letters, 106(25):254101, 2011

    Ernest Montbrió and Diego Pazó. Shear diversity prevents collective synchronization.Physical Review Letters, 106(25):254101, 2011

    2011

  24. [24]

    Motter, Seth A

    Adilson E. Motter, Seth A. Myers, Marian Anghel, and Takashi Nishikawa. Spontaneous synchrony in power-grid networks.Nature Physics, 9(3):191–197, 2013

    2013

  25. [25]

    Phase reduction approach to synchronisation of nonlinear oscillators.Contempo- rary Physics, 57(2):188–214, 2016

    Hiroya Nakao. Phase reduction approach to synchronisation of nonlinear oscillators.Contempo- rary Physics, 57(2):188–214, 2016

    2016

  26. [26]

    Takashi Nishikawa and Adilson E. Motter. Comparative analysis of existing models for power-grid synchronization.New Journal of Physics, 17(1):015012, 2015

    2015

  27. [27]

    Power reservation steering

    NVIDIA. Power reservation steering. NVIDIA Data Center Power Management documentation, https://docs.nvidia.com/datacenter/dps/, 2026

    2026

  28. [28]

    Behind the scenes: GPU power smoothing for large-scale AI training

    Oracle. Behind the scenes: GPU power smoothing for large-scale AI training. Oracle Cloud Infrastructure blog, 9 March 2026, 2026

    2026

  29. [29]

    Antonsen

    Edward Ott and Thomas M. Antonsen. Low dimensional behavior of large systems of globally coupled oscillators.Chaos, 18(3):037113, 2008

    2008

  30. [30]

    Antonsen

    Edward Ott and Thomas M. Antonsen. Long time evolution of phase oscillator systems.Chaos, 19(2):023117, 2009

    2009

  31. [31]

    Thermodynamic limit of the first-order phase transition in the Kuramoto model

    Diego Pazó. Thermodynamic limit of the first-order phase transition in the Kuramoto model. Physical Review E, 72(4):046211, 2005

    2005

  32. [32]

    Cambridge University Press, 2001

    Arkady Pikovsky, Michael Rosenblum, and Jürgen Kurths.Synchronization: A Universal Concept in Nonlinear Sciences. Cambridge University Press, 2001

    2001

  33. [33]

    Rodrigues, Thomas K

    Francisco A. Rodrigues, Thomas K. DM. Peron, Peng Ji, and Jürgen Kurths. The Kuramoto model in complex networks.Physics Reports, 610:1–98, 2016

    2016

  34. [34]

    Self-organized synchronization in decentralized power grids.Physical Review Letters, 109(6):064101, 2012

    Martin Rohden, Andreas Sorge, Marc Timme, and Dirk Witthaut. Self-organized synchronization in decentralized power grids.Physical Review Letters, 109(6):064101, 2012

    2012

  35. [35]

    A soluble active rotator model showing phase transitions via mutual entrainment.Progress of Theoretical Physics, 76(3):576–581, 1986

    Hidetsugu Sakaguchi and Yoshiki Kuramoto. A soluble active rotator model showing phase transitions via mutual entrainment.Progress of Theoretical Physics, 76(3):576–581, 1986

    1986

  36. [36]

    Sanders, Ferdinand Verhulst, and James Murdock.Averaging Methods in Nonlinear Dynamical Systems, volume 59 ofApplied Mathematical Sciences

    Jan A. Sanders, Ferdinand Verhulst, and James Murdock.Averaging Methods in Nonlinear Dynamical Systems, volume 59 ofApplied Mathematical Sciences. Springer, 2nd edition, 2007

    2007

  37. [37]

    Low-frequency oscillations in coupled phase oscillators with inertia.Scientific Reports, 9:17414, 2019

    Huihui Song, Xuewei Zhang, Jinjie Wu, and Yanbin Qu. Low-frequency oscillations in coupled phase oscillators with inertia.Scientific Reports, 9:17414, 2019

    2019

  38. [38]

    Strogatz

    Steven H. Strogatz. From Kuramoto to Crawford: exploring the onset of synchronization in populations of coupled oscillators.Physica D: Nonlinear Phenomena, 143(1–4):1–20, 2000

    2000

  39. [39]

    Strogatz and Renato E

    Steven H. Strogatz and Renato E. Mirollo. Stability of incoherence in a population of coupled oscillators.Journal of Statistical Physics, 63(3–4):613–635, 1991

    1991

  40. [40]

    Lichtenberg, and Shin’ichi Oishi

    Hisa-Aki Tanaka, Allan J. Lichtenberg, and Shin’ichi Oishi. First order phase transition resulting from finite inertia in coupled oscillator systems.Physical Review Letters, 78(11):2104–2107, 1997. 41

    1997

  41. [41]

    Forced oscillations in power systems induced by AI workloads in data centers

    Gustavo Valverde, Georgia Pierrou, Jovan Krajacic, and Gabriela Hug. Forced oscillations in power systems induced by AI workloads in data centers. TechRxiv preprint, 2025. DOI: 10.36227/techrxiv.176127304.46362185

    arXiv 2025

  42. [42]

    Arthur T. Winfree. Biological rhythms and the behavior of populations of coupled oscillators. Journal of Theoretical Biology, 16(1):15–42, 1967

    1967

  43. [43]

    Dynamo: Facebook’s data center-wide power management system

    Qiang Wu, Qingyuan Deng, Lakshmi Ganesh, Chang-Hong Hsu, Yun Jin, Sanjeev Kumar, Bin Li, Justin Meza, and Yee Jiun Song. Dynamo: Facebook’s data center-wide power management system. InProceedings of the 43rd ACM/IEEE International Symposium on Computer Architecture (ISCA), pages 469–480, 2016

    2016

  44. [44]

    M. K. Stephen Yeung and Steven H. Strogatz. Time delay in the Kuramoto model of coupled oscillators.Physical Review Letters, 82(3):648–651, 1999. 42

    1999