Beyond-CMOS Artificial Neuron: A simulation-based exploration of the molecular-FET

The recent growth of Artificial Neural Networks fueled the design of numerous Artificial Intelligence (AI) dedicated hardware implementations. High power dissipation, computational complexity, and large area footprints currently limit CMOS based real-time embedded AI applications. In this work, we design and simulate through SPICE, for the first time, an artificial analog neuron based on the molecular Field-Effect Transistor (molFET) technology. MolFETs are described by a circuital model whose physical characteristics are extracted from atomistic simulations. The designed neuron is a single column of a crossbar-like circuit representing a layer of seven parallel neurons. The drain currents sum up in a soma-like circuit - modelled through a comparator - and trigger the output pulses. We demonstrate the advantages of the molFET in terms of area, power, and speed by comparing it with a conventional MOSFET implementation. The results confirm the molecular technology is a promising candidate for accomplishing high neuron throughput capability and massive redundancy, still providing high energy efficiency. The obtained results foster further investigation of molFET technology both at the device and circuit level.