Enhancing workload-dependent voltage scaling for energy-efficient ultra-low-power embedded systems

Ultra-low-power (ULP) chipsets are in higher demand than ever due to the proliferation of ULP embedded systems to support growing applications like the Internet of Things (IoT), wearables and sensor networks. Since ULP systems are also cost constrained, they tend to employ general purpose processors (GPPs) rather than more energy-efficient ASICs, even though they typically run a single application for the lifetime of the system. Prior work showed that it is possible to reduce the operating voltage and thus the power of such systems without reducing the frequency, since the fixed software stack of a system typically only exercises a subset of a processor's paths, and unexercised paths need not meet timing constraints for the system to work correctly. In this context, we find additional scope for power reduction by intelligently optimizing the processor design based on the system's application-specific activity characteristics to allow an even lower safe operating voltage. We demonstrate automated techniques that maximize the applicationspecific voltage reduction for a system, resulting in 35% additional power savings, on average, compared to the application-specific minimum voltage before optimization and 48% total power savings compared to the original design at nominal voltage.