Sustainable and Synergistic Energy Management for Large Model Computing

Abstract

The rapid growth of foundation models, most prominently large language models (LLMs), is turning the electricity consumption of computing infrastructure into a first-order constraint on the artificial intelligence (AI) industry. This review synthesizes recent literature on three levers for reconciling large-model computing with power-system limits: green energy supply, green computing, and computing-energy synergy (CES), i.e., the coordinated operation of computing infrastructure and power systems. Beyond earlier descriptive accounts, the review makes four contributions: (1) comparative assessments that condense the energy-supply options for AI data centers and the principal energy-efficiency techniques, including quantization, pruning, knowledge distillation, batching, dynamic voltage and frequency scaling (DVFS), and scheduling, into single-view tables with their performance, latency, cost, energy, and carbon trade-offs; (2) an explicit introduction of standard sustainability metrics (PUE, CUE, WUE, energy per token, and carbon emissions per workload) as the quantitative backbone for comparing the surveyed approaches; and (3) a substantially deepened treatment of CES, comparing centralized, distributed, and hybrid algorithmic paradigms, formulating a representative joint scheduling problem, and analyzing feasibility under engineering constraints such as renewable-supply uncertainty, storage limits, grid congestion, workload deadlines, latency requirements, and electricity and carbon price volatility; and (4) a conceptual framework, that links supply-side options, the sustainability-metrics backbone, demand-side workload flexibility, and power-system engineering constraints into a single coordination view. Three research gaps are identified: integrating 24/7 carbon-free electricity procurement with flexible workload scheduling; co-scheduling delay-tolerant training with latency-sensitive inference; and embedding power-system constraints and market signals into CES algorithms. The evidence reviewed indicates that green energy lowers the operational emissions of large-model computing but does not by itself deliver continuous carbon-free operation; green computing techniques can substantially reduce energy per unit of computation at accuracy or latency costs; and CES is a promising coordination paradigm whose large-scale feasibility still needs to be demonstrated under realistic engineering constraints.