Dynamics of PSG-Based Nanosecond Protonic Programmable Resistors for Analog Deep Learning

We study the dynamics of a new class of three-terminal programmable resistors based on proton intercalation into a metal oxide channel using PSG as electrolyte for energy-efficient analog deep-learning hardware accelerators. These CMOS- and BEOL-compatible nanoscale devices exhibit excellent characteristics in terms of high operation speed (5 ns/pulse), high energy efficiency (similar to fJ/pulse), many (1000) nonvolatile conductance states across a large dynamic range (10x), linear and symmetric modulation, and high endurance [1]. This work presents the first real-time study of the dynamics of these devices evidencing impulse-like non-volatile conductance modulation with nanosecond-range pulses without any visible equilibration dynamics. Our study also reveals that the bottleneck to channel conductance modulation is proton transfer across the PSG/WO3 interface and not proton transport across the PSG.