Abstract
Embedded devices using highly integrated chips must cope with conflicting constraints, while executing computationally demanding applications under limited energy storage. Automatic control and feedback loops appear to be an effective solution to simultaneously accommodate for performance uncertainties due to the tiny scale gates variability, varying and poorly predictable computing demands and limited energy storage constraints. This paper presents the example of an embedded video decoder controlled by several feedback loops to carry out the trade-off between decoding quality and energy consumption, exploiting the frequency and voltage scaling capabilities of the chip. The inner loop controls the Dynamic Voltage and Frequency Scaling (DVFS) through a fast predictive control strategy. The outer loop computes the scheduling set-points needed by the inner loop to process frames decoding. The feedback loops have been implemented on a stock PC and experimental results are provided. The upcoming generations of embedded devices are integrating more and more multimedia and telecommunication applications, such as PDAs, mobile phones and tablets, thus requiring increasing on-board computing power. On the other hand, they become even more miniaturised (which means with limited energetic resources) while needing increased autonomy. These constraints are clearly conflicting, therefore technological evolutions are needed to improve their power consumption, computational efficiency and fabrication yield. In this paper, we propose the use of feedback control loops as a possible solution. The approach is applied to an embedded video decoding device whose requirements can be categorised in three main points : a) Energy saving requirements : In current CMOS integrated chips, Dynamic Voltage and Frequency Scaling (DVFS) can be used to efficiently manage the energy consumption of a device. It results in simultaneously managing the frequency and the voltage (note that scaling down the voltage increases signals delays along the paths through electronic gates, thus needing to decrease its clock frequency as well, and inversely (Chandrakasan and Brodersen (1995), Flautner et al. (2004), Varma et al. (2003))). In many cases, the only performance requirement is that the tasks meet their deadline. Such cases create opportunities to run the processor at a lower computing level and achieve the same perceived performance while consuming less energy.A closed-loop DVFS approach is hence a good solution for energy saving. Classically, each task is considered independently, running at a constant voltage whose value is set to meet the deadline, and even selecting a small number of voltage levels leads to a drastic energy reduction (Ishihara and * Corresponding author.