Monolithic 3-D Integration of IAZO/Si-Based 2T0C DRAM With RRAM for Monte Carlo Dropconnect Neural Networks

Edge intelligent systems demand high energy efficiency and robustness, driving interest in Bayesian method-based AI algorithms such as Monte Carlo dropconnect neural networks (MCDNNs) for uncertainty quantification. However, in-memory computing (IMC) for MCDNN faces a critical endurance-retention trade-off. In this paper, we propose a 2T0C-1R IMC unit leveraging heterogeneous volatile/non-volatile memory to resolve this. Fabricated via monolithic 3-D (M3D) integration, it combines an oxide semiconductor write transistor for DRAM retention and a silicon read transistor for RRAM operation. The 2T0C-1R unit achieves >1000 s volatile retention and >4-bit non-volatile storage while meeting endurance requirements of random dropconnect. Experimental MCDNN validation shows a 33% calibration performance improvement over traditional neural networks, highlighting the 2T0C-1R array’s potential for reliable, energy-efficient edge AI applications.