Abstract
With the increase of the number of sensors in an Internet-of-Things (IoT) environment, there is a critical need of in-situ signal processing for local intelligence. This work presents a ring-oscillator based neural network (ONN) for sensor signal processing. A series of 36 ring oscillators are coupled serially and the network performs computations by synchronization. The coupling weights are modulated by sensor inputs which in turn affect the stored memory by indication of synchronization time and frequency of the network. This work studies the effects of synchronization time and frequency of the ONN for different supply voltages, coupling weights and number of oscillating nodes. The 36 oscillator ring neural network scheme has been designed and simulated using 130 nm SiGe BiCMOS process. Simulation results show power consumption to be less than 5.6 mW for the entire ONN and pattern recognition based on synchronization frequency and time differences between the training and input patterns.