Ferroelectric tunnel junction (FTJ) based crossbar array is a promising candidate for the implementation of low-power and area-efficient neuro-inspired computing. In this paper, we fabricated and measured a 10 nm thick Hf _0.5 Zr _0.5 O ₂ FTJ with > 100 on/off ratio and multi-state storage. We found out that the current density of this FTJ is too low for sensing circuitry to distinguish reliably and efficiently. To overcome the low current of FTJs, we suggested using an ultra-thin 1 nm FTJ and employing the concept of stacked capacitors from modern DRAM processes. Following this idea, we further projected a 1 nm thick stacked FTJ and ran the crossbar arrays in the 20 nm node, which shows a desirable summed current level on the order of 10 μA. To model a realistic data pattern of on-state and off-state devices in a 1024 × 1024 array, we mapped quantized weights from a fully connected layer of a deep neural network to the memory cells. The current accuracy and delay are evaluated using array-level SPICE simulations with the consideration of interconnect parasitics. The overall results suggest that FTJ crossbar array is of the potential for realizing neuro-inspired computing.

中文翻译：

---

基于铁电隧道结的神经启发式交叉开关阵列设计

基于铁电隧道结（FTJ）的交叉开关阵列是实现低功耗和面积有效的神经启发式计算的有希望的候选者。在本文中，我们制造并测量了10 nm厚的Hf _0.5 Zr _0.5 O ₂FTJ具有> 100的开/关比和多状态存储。我们发现该FTJ的电流密度太低，无法使感应电路可靠而有效地进行区分。为了克服FTJ的低电流，我们建议使用超薄的1 nm FTJ，并采用现代DRAM工艺中的堆叠电容器的概念。按照这个想法，我们进一步投影了一个1 nm厚的堆叠式FTJ，并在20 nm节点中运行了交叉开关阵列，这显示出理想的总电流水平约为10μA。为了对1024×1024阵列中的开状态和关状态设备的真实数据模式进行建模，我们将量化权重从深度神经网络的完全连接层映射到存储单元。考虑到互连寄生因素，可使用阵列级SPICE仿真评估当前的精度和延迟。