Towards Very Low-Power Mobile Terminals Through Optimized Computational Offloading

Energy consumption is a major issue for modern embedded mobile computing platforms, and with new technological developments, such as IoT and Edge/Fog computing, the number of connected embedded mobile computing systems is rapidly increasing. Heterogeneous multi-core CPUs seek to improve the performance of these platforms, with a particular focus on energy efficiency. By using different techniques like DVFS, core mapping, and multi-threading, a substantial improvement in the achievable CPU energy efficiency level for Multi-processor system-on-chip (MPSoC) can be observed. However, controlling only the CPU power dissipation has a limited effect on the overall platform energy consumption. Other components of the platform, including memory, disk, and other peripherals, play an important role in the energy efficiency of the platform and need to be taken into account. The availability of different sleep strategies at various levels of the platform makes the energy efficiency issue even more complex. In this paper, we set the view of energy efficiency at the entire platform level and discuss computation offloading as a mechanism to help in reaching the optimal platform energy-efficient state. As an application, we consider object detection performed on several types of images to define when offloading is beneficial to the platform energy efficiency. We survey the energy efficiency of different neural network algorithms in an embedded environment, with the possibility to perform computation offloading, and discuss the obtained results concerning the level of object recognition accuracy provided by different neural networks.