Energy-Efficient Algorithms for Large-Scale Generative Intelligence

Abstract

The energy consumption of large-scale generative AI systems has emerged as a critical sustainability concern: training a frontier large language model requires an estimated 100-1000 MWh of electricity, and inference at scale for widely-deployed generative systems can consume comparable cumulative energy through continuous operation. The environmental and economic costs of this energy consumption -- measured in carbon emissions, water cooling requirements, and infrastructure cost -- are increasingly recognised as constraints on responsible and equitable AI development. This paper proposes the Energy-Efficient Generative Intelligence (EEGI) framework, an integrated suite of five algorithmic techniques for reducing the energy consumption of large-scale generative AI systems at both training and inference time: adaptive computation skipping (ACS) that dynamically skips transformer layers for inputs below a complexity threshold; speculative decoding with energy-aware draft models (SD-EA) that uses smaller draft models calibrated for energy efficiency rather than accuracy alone; batching optimisation with dynamic power management (BO-DPM) that co- optimises inference batching and GPU power states; token recycling and reuse (TRR) that caches and reuses attention computations for repeated or similar inputs; and quantisation-aware knowledge distillation (QAKD) that produces smaller, quantised student models retaining generation quality. The EEGI framework is evaluated through energy consumption measurement on a production LLM inference cluster (128 A100 GPUs) running LLaMA-2-70B over 30 days. EEGI achieves a 52.4% reduction in inference energy consumption (kWh per 1M tokens; SD = 4.2%) while maintaining 96.8% of baseline generation quality (MMLU + human evaluation). Training-time QAKD reduces energy by 34.6% relative to baseline training at the same model quality level. The study contributes the EEGI specification, an Energy-Quality Index (EQI), and empirical evidence that coordinated energy optimisation achieves substantially greater reduction than any individual technique.