Memory Switching versus Threshold Memory Switching: Finding a Promising Synaptic Device for Brain-Inspired Artificial Learning Systems

The integration of a selector layer with resistive switching devices has emerged as a promising strategy for developing large-scale cross-point memory by mitigating sneak path currents. However, their performance benefits in obtaining tunable states for emulating the synapses have remained unexplored. In this context, we investigate the device-to-cross-point array (CPA)-level performance of the NbO2-HfOx-based threshold selector-memory switching (TS-MS) device and explore the performance advantages over the HfOx-based memory switching (MS) device for artificial synapses using a fully calibrated multiscale modeling framework. Our findings reveal that the TS-MS device offers highly linear and symmetric long-term potentiation (LTP) and long-term depression (LTD) over the HfOx-based MS device. The NbO2-HfOx-based TS-MS device demonstrates more linear conductance modulation and well-separated multilevel-state operations, which result in a 1.7x reduction in reading inaccuracy and a 4.6x improvement in power efficiency (PE) compared to the MS device, particularly in a 64 x 64 cross-point array under worst-case scenarios. Furthermore, the application of the TS-MS-based bioinspired learning system, with a 15 x 6 cross-point array (CPA), reveals enhanced recognition accuracy and power efficiency over the MS-based cell for a 5 x 3 pixel grayscale image, even in the presence of high noise percentages and intercell wire resistance. Notably, the TS-MS-based CPA demonstrates around 3x reduction in average energy consumption compared to the MS-based CPA for recognizing digital digits. The comprehensive analysis presented in this study suggests that the TS-MS device stands out as a more viable candidate for hardware implementations of brain-inspired artificial learning systems.