Optimizing Power Efficiency: Converting Recurrent Neural Networks to Spiking Neural Networks for Time-Domain Analysis

The deep learning method has recently been used in the modeling of high-speed links to speed up the transient analysis procedure in the time domain. However, neural networks cost a lot of power. Spiking Neural Networks (SNNs) have garnered significant academic interest owing to their distinct characteristic of low power consumption. In this paper, we collected data from circuit to train an RNN for accelerating the transient analysis procedure, then we used an ANN-SNN conversion method to get an SNN. This paper effectively addresses the issue of precision degradation caused during the normalization. By retraining the normalized recurrent neural network (RNN) to map the RNN rules, its predictive performance was restored to that of the RNN model prior to normalization. Our SNN model demonstrated excellent predictive performance while achieving a 294.6 times reduction in power consumption compared to the RNN when test on a Lynxi chip HP201.