Study Finds an Accuracy-Efficiency Paradox in On-Device Energy Forecasting

Summary

A new arXiv paper identifies an “Accuracy-Efficiency Paradox” in on-device energy forecasting. Highly accurate models can consume enough inference energy and accelerate battery aging enough to create a net energy deficit. The authors propose a Total Cost of Ownership framework that combines immediate inference power with the loss of future battery capacity. In thermally sensitive, mission-critical edge environments, the energy saved by complex architectures may be outweighed by their operational intensity. The study recommends optimizing total lifecycle energy loss rather than forecasting accuracy alone.