Constructing an Accurate and a High-Performance Power Profiler for Embedded Systems and Smartphones

The main objective of this paper is to present a new accurate power profiler for embedded systems and smartphones. The second objective is, for it, to be a tutorial explaining the main steps to build power profilers for embedded and mobile systems, in general. We start our work by firstly describing the general methodology of building a power profiler. Then, we showcase how each step is undertaken to build a profiler with two power models. The first one was an artificial neural network (called N-2) that presented a lot of noise in its estimation. After debugging and improvement, the second model, a NARX neural network (we call N-3) was built. It eliminated all the drawback of the first model and had a mean absolute percentage error of 2.8%.