12.1 A 148nW General-Purpose Event-Driven Intelligent Wake-Up Chip for AIoT Devices Using Asynchronous Spike-Based Feature Extractor and Convolutional Neural Network

摘要

Power is a major bottleneck in AIoT devices, which usually operate in random-sparseevent (RSE) scenarios [1] (Fig. 12.1.1, bottom). To process RSEs energy-efficiently, one can integrate an always-on wake-up chip [1] –[5] that only turns on the power-hungry high-performance system (HPS) once valid events are detected. Previous dedicated voice-activity-detection (VAD) wake-up chips using analog front-ends (AFE) followed by neural-network (NN) circuits consume $1 mu mathrm{W}$ [2] and 142nW [3]. Keyword spotting (KWS) wake-up, which is more complicated than VAD, can be achieved by an NN engine at 510nW [4], but it needs an off-chip power-hungry 16b ADC to drive its digital input. For general purpose, an event-driven wake-up chip [1] consumes 57nW. But without an intelligent inference engine (IIE), false detection due to invalid events will consume excessive power. The IIE can be realized by a finite-state-machine (FSM) pattern recognition circuit but with massive $2.2 mu mathrm{W}$ [5]. This work, to our best knowledge, presents a first reported event-driven intelligent wake-up chip that achieves ultra-low power and intelligent event detection for various AIoT RSEs.