Volatile and Non-Volatile Memristive Technologies Implementing Tunable Dynamic and Relaxation for Spiking Neural Networks

Hardware spiking neural networks (SNNs) are currently attracting an increasing interest towards low-power computing systems for edge applications, such as the ones relying on smart analysis of sensory signals in real time. Taking inspiration from brain computation, which processes information with temporal structure, it is desired that also neuromorphic chips compute with dynamics, encompassing devices and circuits able to compute and elaborate information over multiple time scales. In current CMOS-based neuromorphic chips, the integrative functions and temporal dynamics are mainly implemented by capacitors. Anyway, they hardly provide in the system time constants over ms or tens of ms, and their capacitance density is difficult to be scaled-up without occupying large silicon area. Within this framework, memristive technologies have recently shown great promise to substitute or complement CMOS circuits in hardware SNNs, with key role in providing novel and compact dynamical elements which can mimic the neuronal and synaptic features, and are able to extend the time domain of the computing system.<br/>Our work focuses on memristive technologies capable of exhibiting tunable dynamics, integrative functions and relaxation effects which can span up to seconds or longer, up to the non-volatile behavior. The devices are based on the structure Ag/SiO_x/Pt or Ag/Al₂O₃/SiO_x/Pt stacks and the switching mechanism relies on the formation and dissolution of a silver-based filament shorting the two electrodes. The filament can be unstable and exhibits a relaxation behavior once the programming voltage is removed. In the ON, low resistance state, it is also possible to achieve conductance modulation as a function of the input stimuli. We will show that the devices show dynamic behaviors over multiple time scales which originate from an interplay of accumulation and relaxation effects, and the overall features can be controlled by programming conditions and material engineering.