CarbonEdge: Carbon-Aware Deep Learning Inference Framework for Sustainable Edge Computing

A research team has introduced CarbonEdge, a novel framework designed to tackle the growing environmental cost of running AI at the network's edge. While current edge computing systems optimize for speed and throughput, they largely ignore the carbon emissions from deep learning inference workloads. CarbonEdge addresses this by extending adaptive model partitioning—a technique that splits AI tasks across devices—with real-time carbon footprint estimation and a new green scheduling algorithm. This allows the system to make intelligent decisions about where and how to process data, balancing performance against environmental impact through a tunable trade-off.

Experimental evaluations in Docker-simulated, heterogeneous edge environments demonstrate CarbonEdge's significant potential. When set to its 'Green' mode, the framework achieved a 22.9% reduction in carbon emissions compared to standard monolithic execution, where a model runs entirely on one device. More impressively, it boosted carbon efficiency by 1.3x, delivering 245.8 inferences per gram of CO2 versus a baseline of 189.5. This substantial gain comes with negligible operational overhead, adding only 0.03ms of scheduling time per task. The results provide a concrete tool for developers and companies to quantify and actively reduce the environmental footprint of their distributed AI applications, moving the industry toward more sustainable practices.