Energy-Efficient Heterogeneous Computing via Normalized Performance based Proactive Boost for Embedded Artificial Intelligence

Artificial intelligence enabled services are increasingly required for embedded systems. Accordingly, their computing power demand is exponentially increasing. To deal with that, one of the essential technologies is heterogeneous computing. However, the heterogeneous computing requires the excessive amount of energy consumption, since it extensively exploits all available computing units. Thus, it is crucial to reduce the energy consumption in energy-constrained embedded systems. Numerous dynamic voltage frequency scaling mechanisms have been widely investigated to reduce the huge amount of energy consumption. However, they cannot satisfy the desired time of neural network models. To address this limitation, we propose a normalized performance based proactive boost for heterogeneous computing. It dynamically lowers the operating frequencies of all computing units, satisfying the desired time of all neural network models at run-time. We implement our approach into Exynos 2200. We validate the effectiveness of the proposed mechanism through extensive experiments. Our experiments show that the proposed mechanism reduces the energy consumption of all computing units by up to 15.2 percent when compared to that of our earlier work. In addition, our approach incurs negligible run-time overhead.