A Programmable Calculation Unit Employing Memcapacitor-based Neuromorphic Circuit

An efficient calculation unit is developed on the basis of programmable neuromorphic circuit for implementing arbitrary mathematical functions. To retrieve any specific function approximately, the regression algorithm through neural network (NN) is realized by a compact analog circuitry. The memcapacitor technology is associated to Neuron-MOS structure, which couples multiple memcapacitors on the floating gate of a MOS transistor to achieve multiply-accumulation (MAC) operation. In this manner, each synapse of NN is emulated by only one memcapacitor device and the weight is post-fabrication programmable due to the memcapacitive characteristics of memcapacitor. Two types of neurons including sigmoid and linear are implemented by differential pairs with Neuron-MOS transistors. For proof-of-concept, the memcapacitor-based approximate calculation unit (MC-ACU) for arbitrary two-operand computations is achieved with 461 devices. From the real circuit simulation results, all the example functions are retrieved with the maximum inaccuracy of 7.8%. The energy- and device-count of MC-ACU are only 36.7% and 15.7% of that of state-of-the-art works in approximate computing, respectively.