Energy-Constrained AI Inference and Legacy Hardware Optimization for Sustainable ICT Energy Footprint Management

Abstract

The rapid growth of generative AI has disrupted the conventional energy trends of data centers, resulting in sudden energy spikes and accelerating hardware replacement cycles. Although the industry's standard approach involves instant updates for new accelerators, such an approach results in essential environmental oversight because of the failure to consider manufacturing emissions. In this paper, we propose an energy-aware, software-oriented architecture to limit the energy consumption of local infrastructure and prolong the lifetime of legacy enterprise hardware. With the help of request-based model cascades and an emergency load-based admission control module, we are able to decrease the complexity of the models and the accuracy of their inference at times of increased loads. The efficiency of the proposed system is demonstrated through trace-driven numerical experiments using the real-world Alibaba Cluster Trace. Moreover, life cycle optimization modeling shows that retaining legacy hardware yields lower net lifecycle carbon emissions than premature replacement does on grids with abundant renewables.

Keywords

Graphical Abstract

Novelty Statement

Unlike conventional data center upgrades that ignore manufacturing emissions, this paper introduces a novel software architecture that dynamically lowers model complexity and accuracy during peak loads using request-based cascades and emergency admission control. Its primary novelty lies in proving that retaining legacy enterprise hardware on renewable-heavy grids yields significantly lower net lifecycle carbon emissions than the industry-standard practice of immediate, premature accelerator upgrades.