Publication date: 10th April 2024
Memristive devices are highly promising for neuromorphic computing owing to the significant latency and energy consumption improvement potential compared to existing hardware solutions. Their ability to be programmed to multiple resistance states can facilitate the realization of matrix-vector multiply operations within one step by exploiting Kirchhoff’s circuit laws. The accumulative property of the resistance of memristive devices can be utilized to implement artificial neuron network training and can be exploited for realizing local learning rules in spiking neural networks.
Certain memristive material stacks have been reported to exhibit analog switching at least in a limited dynamic range. In this talk, we discuss the relation between the switching dynamics of redox-based memristive devices and their analog programming capability. Theoretical considerations suggest that analog switching can be achieved if a thermal runaway is avoided. Moreover, we argue that internal series resistances play a crucial role in controlling the runaway and determining the accessible resistance window. Experimentally, we compare three groups of devices based on the valence change mechanism with characteristic properties, namely devices with small filaments (HfO2-TiOx) large filaments (SrTiO3) and area-dependent devices (Al/Al2O3/Pr07Ca03MnO3). We will explain the underlying switching mechanism and present the switching kinetics and discuss the possibility for linear and analog resistance programming.
Based on theoretical considerations and experimental analysis we identify general pathways for engineering gradual switching in valence change memories: (i) applying pulses shorter than the transition time, (ii) suppressing thermal runaway by systems showing larger filaments, (iii) including series resistances (ideally nonlinear) for suppressing the thermal runaway, (iv) using area-switching systems showing no Joule heating.
Funded by the German Science foundation within the SFB 917 “Nanoswitches” and by the Federal Ministry of Education and Research project NEUROTEC (Grants No. 16ME0398K and No. 16ME0399) and NeuroSys(03ZU1106AB).