Excellent Synaptic Characteristics and Half-Bias Selectivity in Vertical Short-Channel ECRAM and Selector-Free 4k Cross-Point Array Demonstration

We present a CMOS-compatible, vertical electro-chemical random-access memory (V-ECRAM) designed for the fully-parallel matrix computations for analog AI accelerators. Our V-ECRAM features superior synaptic characteristics, including linear conductance updates, a large on/off ratio (> 17,000), high-speed switching modulation (500 ns), and robust endurance (> 10 M) thanks to the vertical device structure with an increased high-field update area in the channel. Upon confirmation of excellent switching characteristics, we fabricated a 4k V-ECRAM cross-point array with 88.8% yield and evaluated the array-level device statistics and training performance. The V-ECRAM exhibits extremely low cycle-to-cycle variation (0.023), device-to-device variation (0.098), and vector matrix multiplication error (0.18%). Moreover, training capability of the V-ECRAM cross-point array were verified through the hardware implementation of novel Tiki-Taka training algorithms using the improved half-bias selectivity.