Parallel Product-Sum Operation Neuromorphic Systems with 4-bit Ferroelectric FET Synapses

This paper proposes neuromorphic systems using ferroelectric-FETs (FeFET). Since neuromorphic systems mimic neurons, the synaptic weight is stored in non-volatile memory. FeFET is a kind of non-volatile memory and has a feature that the dynamic range of cell current is wide because of FEF characteristics. In order to use the wide dynamic range, proposed neuromorphic systems perform the product of pre-charged voltage as an input and threshold voltage (VTH) of FeFET as the synaptic weight. In addition, proposed neuromorphic systems sum the number of clock pulses. While the synaptic weight can be high precision by using FeFET, the product-sum operation has a problem of low speed. Therefore, this paper also proposes a parallel product-sum operation technique. Proposed neuromorphic systems achieve a large number of product-sum operations at one time. As a result, the total operation time of proposed neuromorphic systems is equivalent to that of conventional neuromorphic systems. In addition, proposed product-sum operation shows 98% lower power consumption than conventional current adder operation.