Nonvolatile Capacitive Crossbar Array for In‐Memory Computing

Abstract

Conventional resistive crossbar array for in‐memory computing suffers from high static current/power, serious IR drop, and sneak paths. In contrast, the “capacitive” crossbar array that harnesses transient current and charge transfer is gaining attention as it 1) only consumes dynamic power, 2) has no DC sneak paths and avoids severe IR drop (thus, selector‐free), and 3) can be fabricated on top of complementary metal–oxide–semiconductor (CMOS) circuits for 3D‐stacking. For the first time, ferroelectric Hf0.5Zr0.5O2 (HZO) capacitive crossbar arrays are experimentally demonstrated. Asymmetry of the HZO electrode interfaces leads to small‐signal capacitance on/off ratio >110% that can achieve read‐disturb‐free operation. The vector matrix multiplication (VMM) experiments are conducted on the fabricated capacitive crossbar array, showing a linear weighted sum versus numbers of input or on‐state weight. The array‐level VMM operation could maintain weight pattern reprogramming after 1) thousands of 1 ms/3 V pulses and 2) an extrapolated 10‐year retention at 85 °C. Array‐level circuit simulation at 22 nm node shows the energy consumption of a capacitive crossbar array is 20–200× lower than the resistive crossbar array counterpart. Moreover, analog‐shift‐and‐add circuits are designed for multibit weight summation, achieving 16.6% less area and 26.9% lower energy consumption than digital‐shift‐and‐add circuits.

An Artificial Neuron with a Leaky Fin‐Shaped Field‐Effect Transistor for a Highly Scalable Capacitive Neural Network

“…In addition, to perform training by a weight update, the fixed ceramic capacitor can be replaced with a capacitance‐tunable synaptic device such as a ferroelectric capacitor or a charge‐storing capacitor. [ 7–9 ] It should be noted that the L‐FinFET neuron per se does not demand any capacitor for a reset circuit, unlike the abovementioned 1C‐1T‐based neuron. In contrast, a neuronal capacitor for n 1C‐1T structured neuron is different from a synaptic capacitor for the capacitive crossbar array.…”

“…To overcome this form of power dissipation during MAC operations, a capacitive neural network that uses a capacitive crossbar array instead of a resistive crossbar array has recently been demonstrated. [5][6][7][8][9] Serious static power consumption by the current flow during reading operations can be prevented because the capacitor consumes only dynamic power. Furthermore, the aforementioned I•R drop in the wires is avoidable owing to the open-circuit characteristic of the capacitor.…”

“…Furthermore, the aforementioned I•R drop in the wires is avoidable owing to the open-circuit characteristic of the capacitor. To implement a capacitive neural network of synapses, various devices, including a memcapacitor, [5] a ferroelectric capacitor, [7,9] and a chargetrap capacitor, [8] have been proposed. However, there is a lack of research on neurons for capacitive neural networks, although they are another important component of a neural network.…”

THz Thin Film Varactor Based on Integrated Ferroelectric HfZrO2

“…3 The C−V characteristics also show an imprint behavior, similar to that observed in Figure 4(a), which is not fully removed upon cycling. The origin of the imprint is most likely due to the accumulation of space charges on one of the capacitor's electrodes, which could increase the pinching of the domain walls 19 and create a Schottky effect. 20 In general, the space charges and defects have a significant influence on the C−V characteristics of thin film ferroelectrics, 20 which could explain why in comparison to P−V characteristics the small-signal C− V characteristics show a more profound imprint.…”

“…Considering the imprint behavior, these ferroelectric varactors can potentially be used in nonvolatile capacitive crossbar arrays for in-memory computing. 19 …”