PaperlayerView exact revision on arXiv

Source-linked AI summary

Low-Latency MoE Inference in Space Data Centers: A Wireless-Power-Transfer Perspective

Haotian Zheng, Zhanwei Wang, Kaibin Huang

arXiv:2610.30923v1cs.DCcs.AIcs.NI

TL;DR

Skewed expert activation mismatches uniform solar-power supply in SpaceDCs, constraining active experts and increasing inference latency. SpaceWPT uses two-sided water-filling control to redistribute power, and experiments report latency reductions of up to 29.25% over No-WPT.

  • Problem

    Skewed expert activation creates heterogeneous power demands despite uniform solar-power supply, constraining high-demand satellites and increasing MoE inference latency.

    Source: §I · p. 1

  • Method

    SpaceWPT minimizes a latency surrogate to derive two-sided water-filling controls that assign satellites to donate, retain, or receive power using activation statistics, workload, and link efficiencies.

    Source: §I · p. 2

  • Results

    29.25% gain over No-WPT for Phi as ηmax increases from 0.35 to 0.75; experiments with Mixtral-8x7B and Phi-3.5-MoE on HumanEval report substantial latency reductions.

    Source: §I · p. 2, C. Effects of System Parameters · p. 11

  • Takeaways & Limitations

    The reported latency gain increases with more compact formations and higher WPT efficiencies.

    Source: §I · p. 2, C. Effects of System Parameters · p. 11

Abstract

from arXiv · show

Space data centers (SpaceDCs) are emerging as platforms for large language models (LLMs), supported by continuous solar-energy harvesting and high-speed interconnections. The mixture-of-experts (MoE) architecture supports distributed LLM inference by deploying experts across satellites. However, skewed expert activation creates uneven computation power demands despite uniform solar-power supply at the dawn-dusk sun-synchronous orbit (SSO). Consequently, active experts remain power-constrained while harvested power at idle experts is underused. These reduce inference efficiency and increase latency. To address this supply-demand mismatch, we propose the SpaceWPT framework, which uses inter-satellite laser-based wireless power transfer (WPT) to redistribute harvested power. The key design challenge is to balance the computation demands of power-donating and power-receiving satellites under WPT losses: transferring more power accelerates computation at the receiver but leaves less power for the donor. After deriving design principles that simplify WPT control, we characterize the resulting latency tradeoff through a tractable surrogate of average inference latency. Minimizing the surrogate yields a novel two-sided water-filling strategy that determines whether each active expert donates, retains, or receives power according to the water level and its WPT link efficiency. Simulations with Mixtral-8x7B and Phi-3.5-MoE show up to 29.25% lower inference latency than without WPT, with larger gains under more compact formations and higher transfer efficiencies.

I. INTRODUCTION

SpaceDCs combine solar-powered satellite fleets with inter-satellite links for distributed LLM inference, but token-dependent expert activation mismatches local power supply and demand. SpaceWPT addresses this challenge by reallocating power among satellites, with a two-sided water-filling control strategy and experiments showing lower latency than without WPT.

I. INTRODUCTION

  • Skewed expert activation mismatches uniform solar supply with heterogeneous power demand, constraining active experts and increasing inference latency.Network-level WPT must choose transfer directions and power levels to balance computation needs at both ends under link losses.Source: §I · pp. 1–2
  • Model-placement and expert-routing approaches redistribute workload, whereas WPT reallocates supply without changing placement or routing.The paper states this avoids expert-transfer overhead and preserves inference accuracy.Source: §I · p. 1
  • Existing laser-based space-WPT studies focus mainly on single-link design, while distributed MoE requires coordinated control across satellites with changing activation-based demands.Space-scale distances and LLM power demands favor laser-based over microwave-based WPT.Source: §I · pp. 1–2
  • SpaceWPT reallocates harvested power among satellites; its near-optimal two-sided water-filling strategy lets each satellite donate, retain, or receive power.A scalar water-level equation determines transfer direction and power according to activation statistics, workload, and link efficiency.Source: §I · p. 2
  • Experiments with Mixtral-8x7B and Phi-3.5-MoE on HumanEval show substantial latency reductions for SpaceWPT over No-WPT baselines.The reported gain increases with WPT efficiency and more compact formations.Source: §I · p. 2

1) MoE Architecture and Inference:

  • Each MoE decoder layer applies shared attention, gating to select the top-κ experts, parallel expert computation, and gateway aggregation.The layer contains K expert feed-forward networks.Source: 1) MoE Architecture and Inference · p. 2

2) Expert Activation Model:

  • The expert activation model uses probability-proportional-to-size sampling without replacement, assigning each expert an importance weight.An expert’s activation probability increases monotonically with its weight.Source: 2) Expert Activation Model · p. 3

3) Computation Latency Model:

  • The computation-latency model uses dynamic voltage and frequency scaling to relate allocated compute power to processor speed and workload latency.The model defines f in FLOPs per second and γ as an architecture-dependent coefficient.Source: 3) Computation Latency Model · p. 3

B. Modelling Space Data Center

The model assigns attention and gating to a gateway satellite and one expert FFN to each of the other satellites. It describes their relative orbital geometry and the resulting gateway–expert communication links.

1) Expert-to-Satellite Mapping:

  • Each of the K expert satellites hosts one expert FFN, while the gateway satellite hosts the attention and gating modules.The gateway is indexed 0, and expert satellites are indexed k ∈ {1, …, K}.Source: 1) Expert-to-Satellite Mapping · p. 3

2) Geometry of Space Data Center:

  • The satellites are modeled in a dawn–dusk sun-synchronous orbit with near-constant illumination and constant harvested solar-power budgets.The local-vertical/local-horizontal coordinate system describes relative motion; the experts follow closed 2:1 elliptical orbits, and distances are treated as fixed during an inference pass.Source: 2) Geometry of Space Data Center · p. 3

C. Communication Model

  • Gateway–expert links use free-space optical communication, with distance-dependent channel gains and latency comprising transmission and propagation delay.Different link distances produce heterogeneous channel gains.Source: C. Communication Model · pp. 3–4

D. Modelling Wireless-Power-Transfer

SpaceWPT uses laser-based inter-satellite power transfer in a gateway-centered star topology. The model specifies how power transfer and token-processing operations affect compute power and inference latency.

1) Laser-Based WPT Model:

  • Each satellite uses a laser transmitter and photovoltaic receiver; end-to-end electrical transfer efficiency combines laser, optical-channel, and receiver conversion efficiencies.The received electrical power is η(d)P, with link efficiency varying by distance.Source: 1) Laser-Based WPT Model · p. 4

2) WPT Topology:

  • In the star topology, each expert exchanges power only with the gateway, and each node’s compute power reflects local solar power plus received power minus transmitted power.Transmit power is nonnegative and capped by the local solar-power budget.Source: 2) WPT Topology · p. 4

E. System Operations

  • Each inference round gates tokens, configures WPT, transmits activations, computes expert outputs in parallel, aggregates them, and computes attention.Configured WPT powers remain fixed during the round’s processing steps, and separate optical links allow WPT during data communication.Source: E. System Operations · p. 4

1) Inference Accuracy:

  • WPT power allocation changes available compute power but does not change inference accuracy.The gating matrix and expert FFNs remain fixed after deployment.Source: 1) Inference Accuracy · p. 5

2) Inference Latency:

  • The per-token latency metric includes the parallel expert phase and gateway attention computation, while gating and WPT-control signaling are omitted as negligible.Expert-phase latency combines communication with computation.Source: 2) Inference Latency · p. 5

III. INFERENCE LATENCY ANALYSIS

The analysis formulates average inference latency as a function of WPT powers and derives principles that reduce the control problem. These principles fix idle-state transfers and rule out simultaneous active-state transmission and reception.

A. Latency-Minimizing Problem

  • The latency-minimizing problem has 4K WPT variables and an activation-state expectation without a closed form, motivating a reduced formulation and tractable surrogate.The objective is average latency over random activation states.Source: A. Latency-Minimizing Problem · pp. 5–6

B. Design Principles for Optimal WPT Control

  • An idle expert optimally donates its full solar-power budget to the gateway and receives no power.An idle expert runs no FFN, so its WPT powers affect gateway power but not expert-side latency.Source: B. Design Principles for Optimal WPT… · p. 6
  • An active expert does not transmit to and receive from the gateway simultaneously, allowing its exchange to be represented by one signed power variable.Positive, zero, and negative values represent donation, neutrality, and reception; received power is reduced by transfer loss.Source: B. Design Principles for Optimal WPT… · p. 6

C. Inference Latency under Simplified WPT Control

The simplified control represents each active expert’s exchange with the gateway by a signed power value. This variable determines expert compute power and, through aggregated transfers, gateway compute power and the latency objective.

C. Inference Latency under Simplified WPT Control

  • For an active expert, signed power Pk determines compute power as P_s − Pk, so donation reduces the expert’s available power and reception increases it.The expert-phase latency is expressed through the signed-power vector.Source: C. Inference Latency under Simplified… · p. 6
  • The gateway’s compute power aggregates contributions from all experts, with donors contributing ηkPk and receivers drawing |Pk|/ηk.The resulting gateway compute power determines attention latency.Source: C. Inference Latency under Simplified… · pp. 6–7
  • Combining expert-phase and gateway-side latency expresses per-token latency through the K-dimensional signed-power vector, whose average is minimized subject to transfer-power constraints.The reduced problem applies under the sufficient condition cited in the formulation.Source: C. Inference Latency under Simplified… · p. 7

D. Tractable Surrogate of Inference Latency

The section approximates average per-token latency with a tractable surrogate that balances gateway attention latency against expert latency. This surrogate yields a convex optimization problem for choosing WPT powers.

D. Tractable Surrogate of Inference Latency

  • The surrogate first averages gateway compute power across activations and approximates the slowest-active-expert latency by the average across active experts.It approximates average per-token latency rather than directly averaging gateway latency over random activations.Source: D. Tractable Surrogate of Inference… · p. 7
  • Increasing power transferred to the gateway shortens attention latency but reduces expert compute power and lengthens expert latency.The opposing monotonicities make each power choice a gateway–expert latency tradeoff.Source: D. Tractable Surrogate of Inference… · p. 7
  • The resulting surrogate objective is convex over its feasible set, making the latency-minimizing power-control problem a tractable convex optimization.The section states that optimizing the surrogate rather than the original objective incurs a bounded average-latency loss.Source: D. Tractable Surrogate of Inference… · pp. 7–8

IV. OPTIMAL WPT POWER CONTROL

The optimal WPT control is characterized by a water level that assigns active experts to donation, neutrality, or reception. Its two-sided water-filling structure depends on link efficiency, and all nonzero active-state transfers share one direction.

A. Optimal WPT Power Control

  • The unique optimal active-state power control for the surrogate problem is characterized by a water level.The water level is the unique solution to an equation; the optimal powers also combine active- and idle-state controls.Source: A. Optimal WPT Power Control · pp. 8–9
  • All nonzero optimal active-state transfers have the same direction, so donation and reception cannot coexist across active-state transfers.The result follows because the direction conditions cannot hold simultaneously for a shared water level.Source: A. Optimal WPT Power Control · p. 8
  • The control forms two-sided water filling: experts donate above an efficiency-dependent floor and receive below an efficiency-dependent ceiling.Higher WPT efficiency lowers the donation floor and raises the reception ceiling, increasing transferred power.Source: A. Optimal WPT Power Control · p. 8
  • The reception ceiling and donation floor define a neutral region that expands as link efficiency decreases, making lower-efficiency experts more likely to remain neutral.The section also presents the active-state control as three modes: donation, neutrality, and reception.Source: A. Optimal WPT Power Control · pp. 8–9

B. Analysis of the Optimal Solution

The section bounds the surrogate solution’s optimality gap and reduces computation to solving one water-level equation. It also considers expert-to-gateway-only transfer as a practical restriction.

B. Analysis of the Optimal Solution

  • The optimality gap between the surrogate-based solution and the original problem is bounded by a constant that depends only on system parameters.The solution is evaluated against the original problem after design principles reduce it without compromising optimality.Source: B. Analysis of the Optimal Solution · p. 9
  • The proposed control determines all K powers by solving one water-level equation in O(K log K) operations.Directly solving the original problem involves 4K power variables and up to the activation patterns referenced in the section.Source: B. Analysis of the Optimal Solution · p. 9
  • The practical unidirectional restriction requires expert-to-gateway transfers and reduces onboard WPT hardware requirements and control overhead.The section uses this restriction as a baseline for measuring the benefit of removing the directional constraint.Source: B. Analysis of the Optimal Solution · p. 9

V. EXPERIMENTAL RESULTS

The evaluation configuration uses a gateway with co-orbiting expert satellites and separate data and WPT links. It specifies the satellite power budget and a shared GPU calibration.

1) SpaceDC Configuration:

  • The simulated SpaceDC uses one gateway and K co-orbiting experts whose gateway–expert distance varies from ρ to 2ρ; the default ρ is 50 m.Each scheme is evaluated at four uniformly spaced orbital snapshots.Source: 1) SpaceDC Configuration · p. 10
  • Each gateway–expert pair has separate FSO data and laser-based WPT links, with a default maximum end-to-end WPT efficiency of 0.689.The data rate is set to 100 Gbps, and the star topology centers the dedicated WPT links on the gateway.Source: 1) SpaceDC Configuration · p. 10

2) GPU Configuration:

  • Every satellite has a 69.09 W solar-power budget and the same GPU profile, calibrated using NVIDIA A100 power–frequency measurements at 990 MHz.The calibration is shared by the gateway and expert satellites; other settings remain fixed unless varied.Source: 2) GPU Configuration · p. 10

3) MoE Configuration:

The evaluation compares SpaceWPT with alternative power-control schemes on Mixtral and Phi, then tests how orbital and computing parameters affect latency. Results show the benefits of demand-driven power redistribution, particularly with more efficient WPT links and compact formations.

3) MoE Configuration:

  • The workload uses native Top-2 routing in 32-layer Mixtral-8x7B-Instruct and Phi-3.5-MoE-Instruct, with 164 HumanEval prompts and 128 decoded tokens per prompt.Recorded expert-pair frequencies preserve activation correlations when evaluating latency, although marginal activation probabilities design the controls.Source: 3) MoE Configuration · p. 10

4) Benchmarking Schemes:

  • No-WPT, Fixed-WPT, Uni-WPT, and SpaceWPT share model deployment, routing, orbital snapshots, and communication and computation settings; only WPT directions and transmit powers differ.Fixed-WPT sends 50% of each idle expert’s solar-power budget to the gateway; Uni-WPT restricts flow to experts-to-gateway, while SpaceWPT permits either direction.Source: 4) Benchmarking Schemes · p. 11

B. SpaceWPT Performance

  • 18.94% lower E2E latency for Mixtral: SpaceWPT reaches 4.6238 ms/token versus 5.7041 ms/token without WPT at the default operating point.Uni-WPT reaches 5.2319 ms/token, an 8.28% improvement over No-WPT; SpaceWPT is 11.62% below Uni-WPT.Source: B. SpaceWPT Performance · p. 11
  • Idle-expert donation shortens gateway attention latency; gateway power transfer to active experts then adds 0.1165 ms/token of attention latency but cuts critical-expert latency by 0.7246 ms/token.Idle experts deliver an average 285.62 W to the gateway after WPT loss; active experts remain neutral under Uni-WPT at the default operating point.Source: B. SpaceWPT Performance · p. 11

C. Effects of System Parameters

  • SpaceWPT’s gain over No-WPT falls as orbital amplitude increases, while higher peak WPT efficiency raises it; gains reach 29.25% for Phi at ηmax = 0.75.From ρ = 40 to 90 m, gains decline from 18.95% to 7.35% for Mixtral and from 28.13% to 19.22% for Phi; from ηmax = 0.35 to 0.75, they rise from 8.78% to 20.21% for Mixtral and 18.80% to 29.25% for Phi.Source: Figure 8 · p. 10, Figure 9 · p. 10, C. Effects of System Parameters · p. 11

VI. CONCLUDING REMARKS

  • SpaceWPT adds energy migration to redistribute power according to computation demand under spatially uniform sunlight; future work considers general exchange topologies and integrated optical WPT-communication links.Combining the links could reduce onboard hardware requirements, but would require balancing WPT efficiency against communication performance.Source: §VI · pp. 11–12

VII. APPENDIX

The appendix constructs a latency surrogate using certainty-equivalent approximations and proves that the bidirectional surrogate optimization has a unique solution. Its KKT conditions characterize donor, neutral, and receiver behavior.

A. Proof of Proposition 1

  • The surrogate approximates gateway latency by placing mean gateway compute power inside the latency nonlinearity, and approximates active-expert maximum latency by the expected average across active experts.The gateway term accounts for active experts’ contributions and idle experts’ WPT delivery, weighted by their activation probabilities.Source: A. Proof of Proposition 1 · p. 12

B. Proof of Theorem 1

  • Strict convexity makes the bidirectional surrogate’s KKT point its unique minimizer; donor, neutral, interior-receiver, and capped-receiver conditions combine into the control relation and balance equation.The proof establishes existence from the strictly convex problem and uniqueness of both the water level and power-control solution.Source: B. Proof of Theorem 1 · p. 12

C. Proof of Corollary 2

The proof compares the original average latency with its surrogate and bounds approximation gaps. It also uses simplified-control properties to establish positive gateway power and bound expert compute powers.

C. Proof of Corollary 2

  • Jensen’s inequality bounds certainty-equivalent gateway latency, while averaging active experts bounds their maximum latency; these approximation gaps are combined to compare the original-latency and surrogate minimizers.The proof invokes the surrogate optimizer to derive the comparison with the minimizer of the original average latency.Source: C. Proof of Corollary 2 · pp. 12–13
  • The simplified control ensures positive gateway power in every state, so the reduced problems need no additional per-state power constraints.The proof also bounds expert compute powers at the optimum using directional consistency, threshold conditions, and the receive cap.Source: C. Proof of Corollary 2 · p. 12
  • The approximation analysis bounds gateway-power variation using the WPT-efficiency range and bounds the gap between maximum and average latency for the active experts.The gateway-side bound uses a Taylor approximation; the expert-side bound depends on the number of active experts.Source: C. Proof of Corollary 2 · p. 12

Related papers

Loading 2610.30923v1…