[B01] Ultra-low Power/IoT Architectures

Abstract

Artificial intelligence (AI) and machine learning (ML)-based decision making is proliferating to application spaces with dynamic and evolving inputs such as internet-of-things (IoTs). The need for real-time decision-making in such applications requires the edge devices in IoT networks to possess extit{in situ} intelligence processing capability. Edge intelligence in the networks is critical to avert unpredictable latency of an otherwise cloud-based intelligence processing. Edge intelligence in IoTs also minimizes their energy demand by avoiding raw data transmission and better preserving data privacy by only transmitting actionable information. Meanwhile, due to form-factor and cost constraints and battery-powered operation, the energy budget and computing/storage resources for edge intelligence are very limited in a typical IoT node. Addressing such computational challenges in IoTs, in this chapter, we review an architectural framework for extit{self-powered edge intelligence}. We first review architectural techniques to exploit sensors in IoTs to harvest energy from their environment to sustain local intelligence processing. Next, we discuss architectures that can identify and focus on regions-of-interest (ROI) while exploiting sparsity in input and intelligence models to minimize edge intelligence workload. Finally, we discuss learning-based architectures to reduce power wastage, such as due to leakage power. With a synergistic integration of the above architectural techniques, many IoTs can leverage self-powered edge intelligence to heighten awareness of their application domains.

Publication
Book chapter of ‘Handbook of Computer Architecture, Springer’