FM-CAC: Carbon-Aware Control for Battery-Buffered Edge AI via Time-Series Foundation Models

A team of researchers from academia has proposed a novel framework called FM-CAC (Carbon-Aware Control via Time-Series Foundation Models) to tackle the growing carbon footprint of edge AI. As billions of devices run always-on, real-time compound AI pipelines, their energy consumption becomes a significant unmanaged source of emissions. FM-CAC addresses this by using an on-device battery as an active temporal buffer, fundamentally decoupling when energy is drawn from the grid from when it is consumed by AI workloads. This allows the system to strategically charge the battery during periods of low-carbon energy availability (like when solar or wind power is high) and discharge it to run AI tasks, thereby maximizing the use of green energy.

At each control step, FM-CAC performs a joint optimization across three dimensions: selecting the most efficient software pipeline variant, tuning the hardware operating point (like CPU/GPU frequency), and managing battery charging/discharging actions. To make proactive decisions, it leverages lightweight, edge-friendly Time-Series Foundation Models (TSFMs) for accurate, zero-shot carbon intensity forecasting without needing task-specific retraining. These forecasts are fed into a dynamic programming solver that uses a deferred cost attribution technique to prevent short-sighted battery depletion, ensuring long-term efficiency. In their evaluations, the framework demonstrated a dramatic reduction in carbon emissions by up to 65.6% compared to baseline methods, all while maintaining near-optimal Quality-of-Service (QoS) and inference accuracy for the AI applications.