2025: The Green Choice: Enabling Sustainable Generative AI through Flexible System Co-Design
The exponential growth of Generative AI poses a critical sustainability challenge, driving unprecedented energy consumption and carbon emissions that threaten to destabilize power grids. Current AI services operate on rigid performance tiers, ignoring a crucial opportunity: user willingness to trade small delays in response time for a significantly lower environmental impact. This project introduces a user-guided co-design approach to create sustainable AI. Researchers will first develop EcoGPT, a tool for conducting large-scale user studies to quantify preferences for “green AI.” Building on these insights, they will design and implement FlexLLM, a novel run-time scheduler that leverages this user flexibility to dynamically map AI tasks across diverse hardware — from cutting-edge GPUs to older, more energy-efficient processors. This work will directly reduce AI’s carbon footprint, enable data centers to become flexible partners for the power grid, and provide the first data-driven framework for creating truly sustainable AI services.
Investigator: Udit Gupta, Cornell Engineering