Ultrathin (<10 nm) Electrochemical Random-Access Memory that Overcomes the Tradeoff between Robust Weight Update and Speed in Neuromorphic Systems

Abstract

Electrochemical random-access memory (ECRAM) devices are a promising candidate for neuromorphic computing, as they mimic synaptic functions by modulating conductance through ion migration. However, the use of a thick electrolyte layer (>40 nm) in conventional ECRAMs leads to an unavoidable tradeoff between synaptic weight updates and operating speed. To address this problem, a Cu-based ultrathin ECRAM (UT-ECRAM) that uses a single 5 nm HfO_x active layer and a ≈1.2 nm AlO_x liner is designed. The highly efficient gate-tunable fast Cu-ion transport in the AlO_x/HfO_x UT-ECRAM enables 1) near-ideal linearity in weight updates (0.45) even achieved with a pulse width (t_w) of 50 μs, 2) dynamic multilevel retention of 10⁴ s, and 3) reliable cycling endurance of 10⁴ cycles. A numerical analysis based on device scaling quantitatively reveals that a relatively high concentration of field-driven Cu ions (≈10²⁰ cm⁻³) contributes to each synaptic weight update per gate voltage (V_G) pulse in the UT-ECRAM without becoming deactivated by traversing thicker layers. This improved gate sensitivity can ultimately overcome the linearity and the ratio/speed tradeoff relationships, paving the way for robust neuromorphic synaptic units.