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

TL;DR

The paper tackles the challenge of leveraging data-center GPU resources for grid support by tightly coupling device-level workload control with distribution-level voltage regulation. It introduces a GPU-to-Grid (G2G) framework that uses the batch size of LLM inference as a fast control knob and employs online feedback optimization to balance grid voltage constraints, user latency, and data-center throughput. By modeling GPU power, latency, and throughput as logistic functions of the log batch size and embedding these into a LinDistFlow-based voltage model, the authors derive a gradient-based OFO algorithm that updates batch sizes in real time and maps them to discrete GPU configurations. OpenDSS-based simulations on the IEEE 13-bus feeder demonstrate substantial reductions in voltage violations compared to tap-only regulation, revealing that reducing GPU power can relieve low-voltage events while increasing power can aid high-voltage conditions, thereby enabling effective, hardware-light grid support from data centers.

Abstract

While the rapid expansion of data centers poses challenges for power grids, it also offers new opportunities as potentially flexible loads. Existing power system research often abstracts data centers as aggregate resources, while computer system research primarily focuses on optimizing GPU energy efficiency and largely ignores the grid impacts of optimized GPU power consumption. To bridge this gap, we develop a GPU-to-Grid framework that couples device-level GPU control with power system objectives. We study distribution-level voltage regulation enabled by flexibility in LLM inference, using batch size as a control knob that trades off the voltage impacts of GPU power consumption against inference latency and token throughput. We first formulate this problem as an optimization problem and then realize it as an online feedback optimization controller that leverages measurements from both the power grid and GPU systems. Our key insight is that reducing GPU power consumption alleviates violations of lower voltage limits, while increasing GPU power mitigates violations near upper voltage limits in distribution systems; this runs counter to the common belief that minimizing GPU power consumption is always beneficial to power grids.

Figures (11)

• Figure 1: GPU-to-Grid (G2G) framework for voltage regulation. The aggregated GPU behavior is simulated based on real measurement data in mlenergy-github, and grid feedback is provided by the power system simulator (e.g., OpenDSS opendss).
• Figure 2: Aggregated power trajectories across batch sizes for the Llama 3.1 405B and Qwen3 235B A22B models.
• Figure 3: Fitted relationships between batch size and performance metrics for three models in the Llama 3.1 family llama3-arxiv24.
• Figure 4: Fitted ITL distributions across batch sizes for the Llama 3.1 8B model.
• Figure 5: Synthetic data center power and average ITL with a fixed batch size of 128 for all LLM models.