Exploring Domain-Specific Architectures for Energy-Efficient Wearable Computing

This paper explores the use of domain-specific architectures for energy-efficient and flexible computing of a variety of workloads, including signal processing applications, in wearable devices. As wearable devices become more popular, and with growing consumer demands, these devices are expected to run a wide range of increasingly complex workloads. A general-purpose solution for wearable computing (e.g., microcontrollers and microprocessors) affords high flexibility, wherein a wide range of applications can be run, but offers mediocre performance and may result in high energy and area overheads. On the other end of the computing flexibility spectrum, application-specific integrated circuits (or accelerators) may optimize a specific algorithm, resulting in inflexible computing and under-utilization of computing resources. Domain-specific architectures (DSAs) provide a happy medium of computing flexibility. DSAs focus on doing a few things—i.e., satisfying the computing requirements of a set of domain workloads with execution similarities—extremely well. As such, DSAs maximize resource usage and achieve substantial performance and energy benefits for a variety of applications. In this work, we first analyze wearable workloads to identify their execution patterns, data movement characteristics, execution bottlenecks, and similarities. Thereafter, we explore various DSA design schemes to meet the increasing processing requirements of wearable workloads, within the typically stringent design constraints of wearable devices. We analyze the performance, energy, and area tradeoffs of the different DSA design schemes in comparison to multiple state-of-the-art architectures, and show, through experimental results, that DSAs offer much promise for flexible, low-overhead, and energy-efficient wearable computing.