An 82nW 0.53pJ/SOP Clock-Free Spiking Neural Network with 40µs Latency for AloT Wake-Up Functions Using Ultimate-Event-Driven Bionic Architecture and Computing-in-Memory Technique
Ying Liu, Zhixuan Wang, Wei He, Linxiao Shen, Yihan Zhang, Peiyu Chen, Meng Wu, Hao Zhang, Peng Zhou, Jinguang Liu, Guangyu Sun, Jiayoon Ru, Le Ye, Ru Huang
Abstract
Human brain is a natural ultimate-event-driven (UED) system with low power and real-time response-ability, thanks to the asynchronous propagation and processing of spikes. Power dissipation and latency are major concerns in AloT devices, usually operating in random-sparse-event (RSE) scenarios (Fig. 22.7.1, top). Being event-driven on the system level, an always-on wake-up system (WUS) detects the valid RSEs energy-efficiently and intelligently, and upon detection turns on the power-hungry high-performance system (HPS). Being event-driven on the module level, a prior WUS [1] uses asynchronous feature extraction and synchronous convolutional neural network to detect the RSEs, consuming 148nW-to-1.68µW with 348ms latency. On the circuit level, the Spiking Neural Network (SNN) gives natural event-driven property. However, the prior SNN works did not fully explore this nature. An SNN circuit [2] achieves keyword spotting task at 205nW-to-570nW, but the framing method causes 100ms latency and is not true real-time. The SNN core in [5] uses synchronous digital design, which consumes significant power by the clock tree. The asynchronous-in-global synchronous-in-local [3]–[4] SNN circuits use local clock signals. They need arbiters in each layer to sort the spikes, weakening the parallelism and timing; additionally, the separation of storage and computing consumes more energy for data movement.