Energy-Efficient Digital Signal Processing Hardware Design.

Abstract

As CMOS technology has developed considerably in the last few decades, many SoCs have been implemented across different application areas due to reduced area and power consumption. Digital signal processing (DSP) algorithms are frequently employed in these systems to achieve more accurate operation or faster computation. However, CMOS technology scaling started to slow down recently and relatively large systems consume too much power to rely only on the scaling effect while system power budget such as battery capacity improves slowly. In addition, there exist increasing needs for miniaturized computing systems including sensor nodes that can accomplish similar operations with significantly smaller power budget. Voltage scaling is one of the most promising power saving techniques due to quadratic switching power reduction effect, making it necessary feature for even high-end processors. However, in order to achieve maximum possible energy efficiency, systems should operate in near or sub-threshold regimes where leakage takes significant portion of power. In this dissertation, a few key energy-aware design approaches are described. Considering prominent leakage and larger PVT variability in low operating voltages, multi-level energy saving techniques to be described are applied to key building blocks in DSP applications: architecture study, algorithm-architecture co-optimization, and robust yet low-power memory design. Finally, described approaches are applied to design examples including a visual navigation accelerator, ultra-low power biomedical SoC and face detection/recognition processor, resulting in 2~100 times power savings than state-of-the-art.