The rise of memtransistors for neuromorphic hardware and In-memory computing - ScienceDirect

Highlights

The state-of-the-art memtransistor from material level to chip level is summarized.
The hardware of memtransistor is focused on a simple device structure.
Applications in in-memory computing and neuromorphic computing are outlined.
Current challenges and outlook in memtransistor are discussed.

Abstract

In the impending era of data deluge, escalating costs related to time and energy owing to data movement underscore the need for a departure from traditional systems. The pursuit of in-memory computing that emulates the high-density memory and energy-efficient information processing of the human brain is of paramount significance. At the forefront of this initiative are memtransistors, with an emphasis on those capable of processing multibit digital and analog data and offering the distinctive features of electrostatically tuning memory and learning behaviors at the device level. Herein, a conceptual overview of the materials and device architectures of memtransistors is presented to underscore their pivotal role in in-memory computing. The discussion includes the strategies for achieving their 3D integration and addresses pertinent challenges and opportunities from materials to chip level for advanced memory storage and neuromorphic computing applications.

Graphical Abstract

Introduction

Transistors are individual electronic components within an integrated circuit that form the foundational building blocks of contemporary computer processors. The exponential increase in the number of transistors in typical processing and memory units since the 1960s is referred to as Moore's law [1], [2]. As transistors continue to scale down, enhanced integration within a limited area has led to the storage and processing of substantial amounts of data, resulting in superior performance [3], [4], [5]. This trend has fueled the development of integrated circuits. However, the proliferation of digital data acquisition and communication has underscored the growing need for rapid data processing for timely decisions [6], [7], [8], [9]. This demand is hindered by inherent structural limitations in conventional technology, which should be scaled down. Furthermore, the contemporary machine learning paradigm relies on complementary metal-oxide-semiconductor (CMOS) technology [10], [11], [12], which induces the bidirectional transfer of extensive data between processing and memory units, resulting in substantial energy consumption. Hence, novel electronic devices with enhanced computing efficiency are required to meet the demands of the ever-expanding information technology landscape.

Memristors have emerged as compelling candidates for such a purpose [13], [14], [15], [16]. An electronic device uses a simple two- (2D) or three-dimensional (3D) matrix configuration, often called a crossbar array, to modulate its conductive state based on the resistance at each node. Memristors have the potential to revolutionize parallel computing by enabling the development of highly efficient and energy-saving neuromorphic hardware [17] which mimics synaptic connections in the human brain. This hardware can process information in parallel to align with the key requirement for rapid and efficient data processing in the digital age. Because of these advantages, memristors have garnered considerable interest in advanced computing, particularly for their efficient facilitation of in-memory computing (IMC) architectures [18], [19]. Their inherent storage and information processing capabilities at the same location make them an ideal solution for applications that prioritize data movement minimization. However, the inherent simplicity of memristors as two-terminal devices that lack selectors such as transistors, complicates the acquisition of multibit data, introduces issues such as sneak current, and hampers the modulation of memristive switching behavior, thus restricting the diversity of learning rules in IMC applications [20], [21], [22]. Thus, memristor capabilities should be expanded by introducing supplementary components that serve as selectors to achieve tunable learning rules and emulate biorealistic functionalities.

The successful fabrication of gate-tunable memristors, known as memtransistors, seamlessly combines the characteristics of transistors and memristors [23], [24]. The three-terminal memtransistor incorporates the advantages of memristors, such as their parallel computing capabilities, and fulfills the crucial role of an inherent selector. Integrating non-volatile memory (NVM) with gate tunability in a memtransistor achieves addressability within diverse crossbar array architectures, thereby enabling bioinspired functionalities. However, the diversity of memtransistor types is constrained by mechanisms such as filament or ion/vacancy migration, coupled with the requisite alteration of channel conductance through electrostatic gating. Current memtransistor materials, such as transition metal dichalcogenides (TMDCs), exhibit only a 2D nature. Consequently, a critical examination of the ongoing advancements in memtransistor development is critical to ensure material diversity, ascertain applicability in IMC, and validate weight modulation performance, and is indispensable for gauging the potential progression of artificial intelligence (AI) and neuro-inspired computing.

Herein, we aim to provide a comprehensive understanding of memtransistors, ranging from their fundamental resemblance to neural network forms inspired by biological analogies to the development of computing-enabled neuromorphic hardware rooted in the simple structure of memtransistors (Fig. 1). These devices are inspired by the structures and functions of biological neurons. Within a perceptron, which serves as a foundational building block, inputs have a parallel role to that of dendrites in biological neurons. Perceptrons receive signals, each of which includes the associated weights that determine their impact on the output. Memristors play a central role in this process akin to synapses, facilitating parallel processing and concurrent computations such as matrix–vector multiplications (MVMs) within a crossbar array. Efforts have been made to address memristor limitations by integrating selectors. In contrast, memtransistors inherently include selectors, thus eliminating the need for additional installations. Further advancements include the development of 3D stacked crossbar arrays achieved by stacking multiple layers of crossbar arrays. This innovative approach enables the creation of multilayer perceptrons, thereby facilitating the construction of neural networks with multiple hidden layers. The integration of such neuromorphic hardware with a variety of neural network structures provides opportunities to implement a wide range of artificial neural networks (ANNs), including deep neural networks (DNNs), recurrent neural networks (RNNs), and spiking neural networks (SNNs) (Fig. 1).

First, we elucidate the concepts and principles underlying representative current memories and resistive random access memory (RRAM), underscoring the significance of the emergence of memtransistors in memory technologies. Subsequently, we offer a concise overview of recent developments in memtransistors, focusing on their materials and mechanisms, and conduct a comparative analysis on the characteristics of memtransistors, memristors, and transistors. Thereafter, we meticulously examine recent progress in computing units and memtransistor arrays, comparing them with various memory technologies and assessing their potential for high-density memory and deep-learning applications. Finally, we present a promising energy-efficient architecture that is scalable to the chip level, while delineating the persist challenges and potential opportunities associated with the transition from individual memtransistor devices to full-scale neuromorphic technology.

Section snippets

Concept and characterization of memtransistor

A memtransistor is a device composed of a material that simultaneously exhibits memristor and transistor characteristics (Fig. 2a). Thus, analogous to memristors, memtransistors exhibit synaptic functionalities, which are crucial for neural network dynamics [24]. Memtransistors also exhibit intricate long-term potentiation (LTP), fostering the persistent synaptic strengthening, and long-term depression (LTD), facilitating synaptic weakening for adaptive learning (Fig. 2b). Inherent short-term

Comparison between current and emerging memory

Over the last decades, the von Neumann architecture has yielded unprecedented success in diverse fields, especially in computing and information technology. However, advances in processing units have far outperformed those in memory, and the high latency and energy consumption of data transmissions have emerged as a bottleneck in the von Neumann architecture, known as the “von Neumann bottleneck” (Fig. 3a) [8]. The advent of the non-von Neumann architecture addresses the challenge of separating

Memtransistor classification

The mechanisms associate with memristors can be broadly categorized into six types: ion vacancy migration, filament formation, ferroelectric effects, tunneling, charge trapping, and phase transition (Fig. 5a). Each mechanism arises from changes in the solid-state behavior of materials, including factors such as band structure and structural modifications such as filament rupture and phase transitions. The aforementioned mechanisms influence the operational characteristics, including the

Memtransistor array

To achieve high-density memory-in-a-chip using memtransistor devices with distinct mechanisms, numerous devices must be integrated within a confined space. The degree of integration and configuration shape of memory varies depending on the type of memory employed. Fig. 8 shows a representation of the memory architectures and their respective characteristics. As previously mentioned, a one-transistor-one-capacitor (1T1C) structure is used in DRAM (Fig. 8a) [90]. The memory cells are arranged in

Memtransistor applications

In the current computational paradigm, performance enhancement relies on the utilization of multiply-accumulate units, wherein data retrieved from external memory undergoes multiplication and subsequent accumulation into registers during each calculation, together with the incorporation of cache, high-speed, limited-capacity memory storage units embedded within the CPU [42]. However, these strategies adhere to the traditional von Neumann architecture, which entails continuous time consumption,

Future memtransistor 3D integration

Although large-scale memory arrays such as 2D memtransistor crossbar arrays have revolutionized IMC technology by combining memory and logic functions, the ever-increasing demand for higher memory densities and enhanced processing power at the chip level necessitates a new perspective. Geometrical scaling times, marked by the relentless pursuit of scaling down, were notably aligned with Moore's law until the early 2000s, after which advancements primarily focused on developing processes that

Challenges and outlook

By exploiting their abundant data and precision, memtransistors present compelling solutions for tasks involving extensive data collection and deep learning. Nevertheless, the adoption of memtransistor-based materials for IMC in neural networks is still in its infancy. Although the overarching architectural framework remains undetermined, recent studies have presented the feasibility of incorporating neural networks into memtransistor-based computing [73], [114], [119], [120]. Despite the

CRediT authorship contribution statement

Wooyoung Shim: Conceptualization, Data curation, Visualization, Writing – review & editing. Jongbum Won: Writing – original draft, Visualization, Investigation, Data curation, Conceptualization, Writing – review & editing. Jihong Bae: Writing – review & editing, Writing – original draft, Visualization, Investigation, Data curation, Conceptualization.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

This research was supported by the Creative Materials Discovery Program through the National Research Foundation (NRF) of Korea, funded by the Ministry of Science and ICT (2018M3D1A1058793).

Jihong Bae received his M.S. degree from the department of Materials Science and Engineering, Yonsei University, Republic of Korea (2021). He is currently pursuing a Ph.D degree under the supervision of Prof. W. Shim. His research interests focus on the synthesis of metastable layered compounds, memory devices and neuromorphic computing.

References (155)

M. Jung et al.(2022)

A review on morphotropic phase boundary in fluorite-structure hafnia towards DRAM technology

Nano Converg.

X. Zhu et al.(2019)

Ionic modulation and ionic coupling effects in MoS2 devices for neuromorphic computing

Nat. Mater.

A. Sawa(2008)

Resistive switching in transition metal oxides

Mater. Today

S. Ye et al.(2021)

Ultimate vertical gate-all-around metal–oxide–semiconductor field-effect transistor and its three-dimensional integrated circuits

Mater. Sci. Semicond. Process.

M.M. Waldrop(2016)

The chips are down for Moore’s law

Nature

G.E. Moore, Cramming more components onto integrated circuits, Reprinted from Electronics, volume 38, number 8, April...
S. Salahuddin et al.(2018)

The era of hyper-scaling in electronics

Nat. Electron

S. Wang et al.(2022)

The Road for 2D Semiconductors in the Silicon Age

Adv. Mater.

Y. Shen et al.(2022)

The Trend of 2D Transistors toward Integrated Circuits: Scaling Down and New Mechanisms

Adv. Mater.

A. Sebastian et al.(2020)

Memory devices and applications for in-memory computing

Nat. Nanotechnol.

Cited by (0)

Jongbum Won received his Ph.D. degree from the department of Materials Science and Engineering, Yonsei University, Republic of Korea (2024). His research interests focus on the synthesis of metastable layered compounds, memory devices.

Wooyoung Shim is a professor at Yonsei University in Korea. He is currently an IBS professor and the director at the Center for Multi-dimensional Materials at Yonsei. At Yonsei, he has been interested in the synthesis of a broad range of nanoscale materials and the development of methods of hierarchical synthesis of multiscale materials, together with the development of nanofabrication tools for future nanotechnology-enabled energy generation and related nanoelectronics.