Analog in-memory computing with memristive devices is a promising solution for overcoming energy inefficiencies of traditional Von Neumann architectures, especially in deep learning applications. However, filamentary memristive devices encounter significant challenges, such as high forming and set voltages and limited resistance in analog switching regions, causing excessive power consumption. Here, an engineering strategy is presented that reduces operational voltage by addressing ionic supply bottlenecks and lowers analog switching current using a dual-matrix filamentary switching approach. GeSe2 is utilized as a high-mobility matrix and densified amorphous silicon as a low-mobility matrix, along with Ag and Pt nano-cluster layers for dual-matrix devices. Experimental results show over a 50% reduction in both forming and set voltages and more than a 96% decrease in reset current compared to pristine devices. Moreover, the proposed devices exhibit a 93% reduction in analog energy consumption (0.98 pJ) compared to pristine devices, stable retention for ≈24 h, and endurance for ≈50k cycles. Furthermore, simulations on the Spiking-VGG9 architecture, employing quantization-aware training and the tiki-taka method in the resistive processing unit framework, demonstrate accuracies of 89.38% for CIFAR10 and 63.70% for CIFAR10-DVS while concurrently reducing total energy consumption by 60.42% relative to pristine devices.
Kang et al. (Mon,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: