Sensor Systems and Software. 7th International Conference, S-Cube 2016, Sophia Antipolis, Nice, France, December 1-2, 2016, Revised Selected Papers

Research Article

A Dual Processor Energy-Efficient Platform with Multi-core Accelerator for Smart Sensing

Download
357 downloads
  • @INPROCEEDINGS{10.1007/978-3-319-61563-9_3,
        author={Antonio Pullini and Stefan Mach and Michele Magno and Luca Benini},
        title={A Dual Processor Energy-Efficient Platform with Multi-core Accelerator for Smart Sensing},
        proceedings={Sensor Systems and Software. 7th International Conference, S-Cube 2016, Sophia Antipolis, Nice, France, December 1-2, 2016, Revised Selected Papers},
        proceedings_a={S-CUBE},
        year={2017},
        month={7},
        keywords={Low power design Sensors platform Energy efficiency Power management Multi-core processor},
        doi={10.1007/978-3-319-61563-9_3}
    }
    
  • Antonio Pullini
    Stefan Mach
    Michele Magno
    Luca Benini
    Year: 2017
    A Dual Processor Energy-Efficient Platform with Multi-core Accelerator for Smart Sensing
    S-CUBE
    Springer
    DOI: 10.1007/978-3-319-61563-9_3
Antonio Pullini1,*, Stefan Mach1,*, Michele Magno1,*, Luca Benini1,*
  • 1: ETH Zurich
*Contact email: antonio.pullini@iis.ee.ethz.ch, stefan.mach@iis.ee.ethz.ch, michele.magno@iis.ee.ethz.ch, luca.benini@iis.ee.ethz.ch

Abstract

Energy-efficient computing has increasingly come into focus of research and industry over the last decade. Ultra-low-power architectures are a requirement for distributed sensing, wearable electronics, Internet of Things and consumer electronics. In this paper, we present a dual-mode platform that includes an ultra-low power Cortex Arm M4 microcontroller coupled with a highly energy efficient multi-core parallel processor. The platform is designed to maximize the energy efficiency in sensors applications by exploiting the Cortex Arm M4 to achieve ultra-low power processing and power management, and enables the multi-core processor to provide additional computational power for near-sensor data centric processing (i.e. accelerating Convolutional Neural Networks for image classification) increasing energy efficiency. The proposed platform enhances the application scenarios where on-board processing (i.e. without streaming out the sensor data) enables intensive computation to extract complex features. The platform is geared towards applications with limited energy budget, as for example in mobile or wearable scenarios where the devices are supplied by a battery. Experimental results confirm the energy efficiency of the platform, demonstrate the low power consumption, and the benefits of combining the two processing engines. Compared to a pure microcontroller platform we provide a boost of 80× in terms of computational power when running general purpose code and a boost of 560× when performing convolutions. Within a reasonable power budged of 20 mW compatible to battery-operated scenarios the system can perform 345 MOPS of general purpose code or 1.5 GOPS of convolutions.