GreenAuto: An Automated Platform for Sustainable AI Model Design on Edge Devices

TL;DR

GreenAuto addresses the environmental impact of edge AI by presenting an automated platform that combines an energy-focused, expanded NAS search space with Pareto-front optimization guided by gradient descent. It leverages pre-trained kernel-level energy predictors and a NASWOT accuracy proxy to rapidly estimate $E_p$ and $N_s$ across a large space, while on-device measurements validate actual energy use on target hardware. The system demonstrates substantial sustainability gains, achieving competitive accuracy with dramatically lower energy and carbon footprints than traditional NAS approaches (e.g., per-model carbon 0.013 kgCO$_2$ vs 0.231 kgCO$_2$ for NASNet-A) and enabling autonomous, hardware-aware exploration. These results highlight GreenAuto’s potential to enable scalable, low-impact AI deployment on billions of edge devices, with broad applicability across energy-constrained environments.

Abstract

We present GreenAuto, an end-to-end automated platform designed for sustainable AI model exploration, generation, deployment, and evaluation. GreenAuto employs a Pareto front-based search method within an expanded neural architecture search (NAS) space, guided by gradient descent to optimize model exploration. Pre-trained kernel-level energy predictors estimate energy consumption across all models, providing a global view that directs the search toward more sustainable solutions. By automating performance measurements and iteratively refining the search process, GreenAuto demonstrates the efficient identification of sustainable AI models without the need for human intervention.

Figures (3)

• Figure 1: System overview of GreenAuto, operating autonomously without human intervention.
• Figure 2: Timing sync between edge device and external power monitor
• Figure 3: Pareto front update and the best model selection in each iteration.