On the infeasibility of analysing worst-case dynamic energy

This paper examines dynamic energy consumption during the execution of software on deeply embedded IoT microprocessors. In worst-case energy consumption analysis, energy models are used to find the most costly execution path. Few models adequately consider dynamic energy caused by switching due to operand values. We find that energy contribution of operand values can be significant, prove that finding worst-case input data is NP-hard, and that tight bounds cannot be approximated with guaranteed safety. We conclude that accurate worst-case analysis of data-dependent energy is infeasible, and provide recommendations to minimize the impact of this, without resorting to significant over-estimation.