Sustainable Intelligent Systems for Energy-Efficient Computing

Abstract

The computational demands of modern intelligent systems — driven by the scaling of foundation models, the proliferation of edge AI deployments,
and the growth of always-on inference services — are generating an energy consumption trajectory inconsistent with global climate commitments
and the European Green Deal's 2050 net-zero target. Training a single frontier language model can emit hundreds of tonnes of CO2 equivalent, while
global AI inference energy consumption is projected to surpass total EU data centre energy use by 2030. Sustainable intelligent systems —
architectures, algorithms, and deployment strategies that minimise energy consumption and carbon footprint without disproportionate sacrifice of task
performance — are therefore a technical and societal imperative. This study presents GreenAI, a unified sustainable intelligent systems framework
integrating hardware-aware neural architecture search (HW-NAS) for energy-optimised model design, dynamic voltage and frequency scaling
(DVFS)-aware inference scheduling, and carbon-aware workload orchestration across heterogeneous green-powered computing infrastructure.
GreenAI was evaluated across five AI workload classes (image classification, NLP inference, time-series forecasting, reinforcement learning, and
generative image synthesis) on three European green computing testbeds: the GENCI Jean Zay supercomputer (France, 100% renewable), the
Jülich Supercomputing Centre JUWELS Booster (Germany, 100% renewable), and the UP Green HPC cluster (Portugal, 87% renewable). GreenAI
achieved energy efficiency improvements of 47.3% on average over standard deployment baselines, carbon footprint reduction of 52.8%, and task
accuracy retention of 97.4% relative to unconstrained baselines. DVFS-aware scheduling reduced peak power draw by 31.4%. Carbon-aware
orchestration shifted 38.7% of flexible workloads to high-renewable-availability time windows. These results establish GreenAI as a comprehensive,
practically deployable framework for sustainable energy-efficient intelligent computing across European green infrastructure.