Voltage Bias Scheme Optimization in FeFET Based Neural Network System

In this work, we focused on the performance optimization of the neural network (NN) system in the synaptic device of HfAlOx (HAO)-based ferroelectric field-transistors (FeFET), especially with the voltage bias scheme. The weights linearity, power consumption, and current ratio of the HAO-FeFET based synapse are measured under different voltage bias. Using the open-source tool "NeuroSim $^{\mathbf {+{''}}}$ , the training accuracy, and power efficiency can be evaluated. It shows that the performance of the NN is highly dependent on gate/drain's voltages bias of the FeFETs. To suppress the nonlinearity of both the potentiation and depression during online training, the FeFET were biased at the triode and saturation region respectively under a large gate voltage bias. However, the power consumption will increase for inference, which should be a trade-off for system optimization.