Charge-Domain Neuromorphic Computing: from Energy-Efficient Adiabatic Circuits to Ferroelectric Memcapacitors

Artificial Intelligence is revolutionizing our way of living, but its ever-growing energy demand limits its widespread diffusion on the edge. With the aim of enabling a new generation of intelligent edge devices with unprecedented energy-efficiency, this thesis explores the potentials of the neuromorphic computing paradigm in the charge domain from multiple points of view, encompassing circuit design and device-level characterizations. The first adiabatic Leaky-Integrate-and-Fire neuron with a tunable refractory period is presented in detail. The proposed adiabatic neuron is able to emulate the most important primitives for a valuable brain-inspired computation with biologically plausible dynamics, all the while exploiting both the charging and the energy-recovery phases enabled by an LC resonance to ensure a seamless and continuous computation. In particular, extensive simulations unveil an adiabatic efficiency between 90% and 99% over a wide range of resonance frequencies, ultimately resulting in a 9x energy saving with respect to a non-adiabatic driving, and in an energy per synaptic operation orders-of-magnitude lower than other state-of-the-art solutions. The scaling of the design is limited by the large size of the synaptic capacitors, thus making non-volatile memcapacitors, namely devices capable of storing information in their capacitance value, a compelling prospect to complement the adiabatic operation. Thanks to their low-power switching and BEOL compatibility, ferroelectric-based memcapacitors are the most promising technological platform to shift the neuromorphic computation towards the charge domain, so as to achieve better performance in terms of both reliability and energy-efficiency. By making use of a versatile in-house-developed experimental setup for the characterization of ferroelectric tunnel junctions and capacitors, novel methodologies are devised to understand the gap between the large- and small-signal capacitive responses of ferroelectrics, and the comparison with physics-based simulations ultimately identifies in the small-signal capacitance the true memcapacitance of the devices. Finally, a hybrid device capable of operating either as a filamentary memristor or as a ferroelectric memcapacitor has been statistically characterized in terms of multi-level programming and read-disturb, variability and scalability. The strong variability of these memcapacitors currently hampers their adoption as conventional synaptic weights, nonetheless it can be conveniently exploited to introduce heterogeneity in the dynamics of synapses and neurons. In this respect, a recurrent neural network that leverages the same hybrid devices as both memristors and memcapacitors is calibrated against experimental data and trained. The resulting performance are comparable to the state-of-the-art, showcasing the potential of hybrid memristive-memcapacitive neuromorphic systems in emulating bio-inspired heterogenous networks.