GPU-to-Grid: Distribution Voltage Regulation via GPU Utilization Control

Zhirui Liang et al. (arXiv:2602.05116) have submitted a paper presenting a GPU-to-Grid framework that couples device-level GPU control with power system objectives (submitted v1 on 4 Feb 2026; PDF 2,474 KB).

• Main announcement: The paper introduces a GPU-to-Grid framework and an online feedback optimization controller that uses measurements from both the power grid and GPU systems; it proposes using batch size during LLM inference as a control knob to trade off GPU power consumption against inference latency and token throughput (arXiv PDF: https://arxiv.org/pdf/2602.05116, arXiv ID: 2602.05116).
• Background and details: The work formulates the problem as an optimization problem, targets distribution-level voltage regulation, and reports the key insight that reducing GPU power alleviates lower-voltage violations while increasing GPU power can mitigate upper-voltage violations — a result that contradicts the common belief that minimizing GPU power is always grid-beneficial. Subject classification: Systems and Control (eess.SY).