This article presents a 148-nW always-on wake-up system that drastically reduces the system power consumption of Internet of Things (IoT) sensor nodes while oftentimes operating in random-sparse-event (RSE) scenarios. To significantly reduce the long-term average (LTA) power consumption and realize multiapplication and intelligent event detection, three techniques are proposed: 1) In a three-stage pipelined event-driven architecture, a frame generator and a convolutional neural network intelligent inference engine (CNN IIE) in stage III are event-driven by the preliminary detectors in stage II, and the detectors are triggered by a level-crossing (LC) analog-to-digital converter (ADC), i.e., stage I, dramatically reducing the overall power consumption. 2) The clock-free pulse-based instant rate of change (IROC) feature extractor directly processes the asynchronous pulses of the LC-ADC outputs in the temporal domain instead of utilizing a conventional power-hungry frequency-domain method. 3) A reconfigurable IROC, the frame generator, and the CNN IIE provide adaptive intelligence for various IoT events, enhancing the accuracy of multipurpose detection with ultralow power. We demonstrate two artificial intelligence IoT (AIoT) applications at 0.6-V VDD. For electrocardiogram (ECG) recognition, one example works at a typical event rate (ER) of 4800/h, with an active power of 1.68 $μ \textW$ and a precision of up to 99%; the other is used for keyword spotting (KWS), where the chip achieves 378 nW at 720/h ER and 94% accuracy. The LTA power is bounded to 148 nW, while the event-driven chip is on call and waiting for events; this chip dominates the AIoT device battery life in RSE scenarios.
Generally, computer-science-oriented artificial neural networks (ANNs) and neuroscience-oriented spiking neural networks (SNNs) are two main approaches to develop brainspired non von Neumann computing systems. The goal of exploring complex artificial intelligence (AI) systems demandsgeneral neuromorphic hardware platforms compatible with both of them. However, as a result of the obvious differences in their fundamental mathematical expression and coding scheme, many neuromorphic platforms or deep neural network (DNN) accelerators accommodate only one of them. This brief presents a reconfigurable scalable neuromorphic chip based on digital leaky integrate-and-fire (LIF) neuron model targeting low-cost large-scale systems. By unifying ANN and SNN paradigms within a LIF neuron framework with point-to-point (P2P) communication, the chip can accommodate most popular neural networks. The chip adopts distributed on-chip memory architecture with a capacity of 64K neurons and 64M synapses. It achieves a peak throughput of 12.29 GSOP/s at 1.2 V, 192MHz and peak energy efficiency of 2.64 pJ/SOP at 890 mV, 24MHz.The results of implementations of a spike-based spatio-temporal memory model and ternary-weight event-based convolutional neural networks (CNNs) demonstrate outstanding compatibility of the chip.