PREDICTING MICROPROCESSOR POWER CONSUMPTION BASED ON HARDWARE PERFORMANCE COUNTERS

Abstract

This thesis presents a foundation for microprocessor power consumption estimation and prediction framework development for ARM-based devices with limited power resources using hardware performance counters. The study introduces a LSTM RNN model to estimate power consumption based on CPU HPC data without evaluation of other hardware events.. This method has a potential advantage for battery-powered embedded systems, where traditional power measurement tools have small efficiency. The research builds upon previous work in the field, highlighting the importance of energy-efficient designs in the growing IoT market. The proposed framework aims to enhance the battery life of portable devices, by helping developers to optimise the software and enabling devices with real-time power management. The model was trained on the dataset collected in idle, video recording, video streaming and audio recording scenarios and evaluated on RMSE performance. Results of the paper suggest that prediction performance of the RNN LSTM model are lacking, but the use of adaptive algorithms in a regression like evaluation have potential to be effective.