Lifecycle Assessment Framework for Large-Scale AI Deployments

Abstract

The growing success of large-scale Artificial Intelligence (AI) systems, such as GPT-4 and Gemini, has raised worries about the long-term sustainability of these systems from the environmental, economic, and operational perspectives. While these systems have revolutionary potential in analytics, automation, and intelligent decision making, they have a high energy cost, high carbon emission, hardware usage, and infrastructure overhead over their life-cycle. Existing sustainability evaluations tend to focus on single specific metrics such as training energy and do not take a full lifecycle approach. To fill this attention table TEI2 proposes a a Lifecycle Assessment (LCA) based framework, appropriate for massive-scale AI systems. The framework brings together multidimensional numbers such as carbon emissions (kgCO2e), energy (kWh), hardware (%), water (), working costs and computational speed (FLOPs per infer). It takes advantage of process-based LCA (life cycle assessment) modelling, real time cloud telemetry and multi-objective evolutionary optimisation. Experimental evaluation on distributed GPU clusters shows that there are great improvements, including 28.4% energy saving, 32.7% carbon emission reduction, and 18.9% cost saving, while the accuracy of the model remains unchanged. The proposed approach is a low-cost and policy-aligned means of sustainable deployment of AI.