Fusion synapse by memristor and capacitor for spiking neuromorphic systems

Research on neuromorphic computing with spiking neural networks and in-memory computing to achieve low-power consumption and high-speed operation has received great attention. In this study, we proposed a new synaptic device called a fusion synapse that consists of a memristor and a capacitor connected in a series and uses a change in the time constant as a synaptic weight. We also prepared both neuron and synchronizer circuits to utilize the time constants as synaptic weights and trained them with PyTorch according to two models. One model is based on previous research, and the other is an aware model of circuit information. We converted the synaptic weights generated by the training into conductance. When assembling two models for inference, 95% accuracy was achieved on MNIST, and the power consumption per inference was 640 nJ. The accuracy was comparable to related research, and considering Denard scaling, the power consumption was lower than the digital implementation.