1 Transistor-Dynamic Random Access Memory as Synaptic Element for Online Learning

Abstract

The rapid advancements in the field of autonomous systems have led to a significant demand for artificial-intelligence-of-things (AIoT) edge-compatible neuromorphic training accelerators with continual/online learning capability. These accelerators require a large network of synaptic elements with high degree of plasticity, high endurance, large integration density, and ultralow programing energy. Although emerging nonvolatile memories exhibit promising potential as synaptic devices, their widespread application in training accelerators is limited due to their low endurance and immature fabrication technology. In contrast, capacitor-less 1 transistor-dynamic random-access memories (1T-DRAMs) have recently emerged as lucrative alternative to the conventional (1T/1C) DRAMs owing to their high scalability and low footprint. Considering the high endurance, large integration density, and ultralow write energy of the 1T-DRAMs, in this work, for the first time, their potential is explored as synaptic elements for online learning. The proposed 1T-DRAM-based synaptic element exhibits multi-level capability (up to 6 bits), a large dynamic range (3.91 × 10³), an ultralow energy, and an appreciable linearity for potentiation/depression. The 1T-DRAM-based synaptic element also exhibits a paired pulse facilitation with an exponential decay similar to the biological synapses. Furthermore, a multilayer perceptron utilizing the proposed 1T-DRAM synapses achieves an accuracy of 87.10% on MNIST dataset.