Memristive Devices for Brain-Inspired Computing

From Materials, Devices, and Circuits to Applications - Computational Memory, Deep Learning, and Spiking Neural Networks

Sabina Spiga, Abu Sebastian, Damien Querlioz, Bipin Rajendran

• Publisher: Woodhead Publishing
• Year: 2020
• Print ISBN: 9780081027820
• eBook ISBN: 9780081027875
• Topic: Technology & Engineering
• Subtopic: Electrical Engineering & Telecommunications
• Index: Technology & Engineering

Memristive Devices for Brain-Inspired Computing: From Materials, Devices, and Circuits to Applications—Computational Memory, Deep Learning, and Spiking Neural Networks reviews the latest in material and devices engineering for optimizing memristive devices beyond storage applications and toward brain-inspired computing. The book provides readers with an understanding of four key concepts, including materials and device aspects with a view of current materials systems and their remaining barriers, algorithmic aspects comprising basic concepts of neuroscience as well as various computing concepts, the circuits and architectures implementing those algorithms based on memristive technologies, and target applications, including brain-inspired computing, computational memory, and deep learning.This comprehensive book is suitable for an interdisciplinary audience, including materials scientists, physicists, electrical engineers, and computer scientists.- Provides readers an overview of four key concepts in this emerging research topic including materials and device aspects, algorithmic aspects, circuits and architectures and target applications- Covers a broad range of applications, including brain-inspired computing, computational memory, deep learning and spiking neural networks- Includes perspectives from a wide range of disciplines, including materials science, electrical engineering and computing, providing a unique interdisciplinary look at the field

Part I

Memristive devices for brain–inspired computing

Outline

• Chapter 1 Role of resistive memory devices in brain-inspired computing
• Chapter 2 Resistive switching memories
• Chapter 3 Phase-change memory
• Chapter 4 Magnetic and ferroelectric memories
• Chapter 5 Selector devices for emerging memories

Chapter 1

Role of resistive memory devices in brain-inspired computing

Sabina Spiga¹, Abu Sebastian², Damien Querlioz³ and Bipin Rajendran⁴, ¹CNR–IMM, Agrate Brianza, Italy, ²IBM Research – Zurich, Rüschlikon, Switzerland, ³Centre for Nanoscience and Nanotechnology, Universite Paris-Saclay, Palaiseau, France, ⁴Department of Engineering, King’s College London, London, United Kingdom

Abstract

This chapter introduces the current state of the art of memristive devices and their roles in emerging computational paradigms. The materials systems and device characteristics of various resistive switching memories, also named memristive device technologies, including PCM, RRAM, MRAM, and FeRAM, are presented and discussed in terms of functionality, switching mechanism, and perspectives for applications beyond storage, toward brain-inspired computing. Conventional digital computers face increasing difficulties in performance and power efficiency due to their von Neumann architecture. Memristive devices are therefore considered as promising hardware building blocks for data and memory-centric computing schemes. This chapter provides an overview of currently enabled novel applications exploiting device properties.

Keywords

RRAM; PCM; FeRAM; FeFET; MRAM; MTJ; memristive devices; resistive switching memory; in-memory logic; in-memory computing; analog computing; neuromorphic computing; neural networks; deep neural networks; spiking neural networks

1.1 Introduction

This chapter introduces various types of resistive memory devices (also named memristive devices) of current interest for brain-inspired computing. These memristive device technologies include a broad class of two- or three-terminal devices whose resistance can be modified upon electrical stimuli. The resistance changes can last for short- or long-time scales, leading to a volatile or nonvolatile memory effect, respectively. Memristive devices are based on a large variety of physical mechanisms, such as redox reactions and ion migration, phase transitions, spin-polarized tunneling, and ferroelectric polarization. The switching geometry can involve a volume, interfacial, or confined 1D filamentary regions [1–8].

Although these technologies have been mainly developed as nonvolatile memory devices for storage applications, recently, they have been receiving increasing interest for brain-inspired computing, and many exciting developments are underway in this direction [1,9–23]. Today we are facing a revolution driven by the increasing amount of data generated each day, which need to be stored, classified, and processed, leading to the paradigm of data-centric-computing. On the other hand current computing systems are inherently limited in energy efficiency and data bandwidth by the physically separated memory and processing units (von Neumann bottleneck), as well as by the latency mismatch between the memory and processing units (memory wall) [9,10,13]. Memristive devices have the potential to meet the considerable demand for new devices that enable energy-efficient and area-efficient information processing that transcends von Neumann computing. In the following sections we describe the leading memristive technologies and their current potential for various applications.

1.2 Type of resistive memory devices

Fig. 1.1 shows a classification of the most representative and mature resistive switching memory technologies (RRAM, PCM, MRAM, and FeRAM) based on their underlying physical mechanism, location of the switching region, and their current–voltage or resistance–voltage behavior.

Figure 1.1 (A) Sketch of resistive random access memory (RRAM) devices featuring filamentary (top) and interfacial switching (bottom). The corresponding representative current–voltage characteristics are reported on the right, indicating bipolar switching from a high- to low-resistance state (set) and vice versa (reset). For filamentary systems an initial high-voltage forming is required. (B, Top) Schematic drawing of phase-change memory (PCM) cell in the crystalline (LRS) and amorphous state (HRS) of the chalcogenide material. (Bottom) Current–voltage and resistance-programming power characteristics. (C) Ferroelectric random access memory (FeRAM) and its polarization–electric field hysteresis (left), and ferroelectric field effect transistor (FeFET), with corresponding drain current versus gate voltage characteristics (right). (D) Sketch of a magnetic tunnel junction (MTJ) (top) and evolution of the resistance versus applied voltage (bottom) for the parallel and antiparallel configurations of the magnetization orientation of the pinned and free layers. The MTJ is the key element for a magnetic random access memory (MRAM).

Resistive random access memory (RRAM) shown in Fig. 1.1A is based on a two-terminal structure where a switching medium is sandwiched among two electrodes, whose resistance can be switched reversibly among two or more states by applying external electrical stimuli [1–3,24–28]. Many materials can be used for the switching medium, including inorganic and organics ones (transition metal oxides, perovskites, chalcogenides, polymers, etc.), and the resistive switching process typically involves the creation and rupture of conductive filaments (CF) shorting the two electrodes. These types of devices are also named filamentary RRAMs, the low-resistance state (LRS) occurs when a CF bridges the two metal layers (Fig. 1.1A-top), while the high-resistance state (HRS) is achieved by partial dissolution of the filament. The devices based on oxygen ion migration effects and subsequent valence change of the metals in metal oxides are named valence-change memory (VCM) or anion-based memory, and the resulting CF is formed by a localized concentration of defects. Examples of VCM RRAM are related to material systems based on HfO_x, TaO_x, TiO_x, SiO_x, WO_x, Al₂O₃, or other transition metal oxides in combination with TiN, TaN, Ti, Ta, and Pt electrodes [3,14,26,28]. There is another class of devices where the CF is formed by the metal ion movement from the active electrode (often Ag or Cu) into the switching medium, which are named as electrochemical metallization memory (ECM), cation-based memory, or even Conductive Bridge Random Access Memories (CBRAM) [4,14,26,27]. These types of RRAMs are fabricated by using an active electrode (Ag, Cu, Co, etc.) in combination with an inert electrode (TaN, W, Pt, etc.), while the switching medium is based on solid electrolytes such as GeSe, GeS, Ag_xS, Cu_xS, or even oxides such as SiO₂. For filamentary RRAMs, VCM or CBRAM, an initial electroforming step is necessary to establish the first filament formation; after that forming step, the cell can be repeatedly switched between the LRS and HRS (cycling endurance) up to 10⁴–10¹² cycles, depending on the materials systems and if the measurements are done at a single device or array level. RRAM is considered a nonvolatile memory since the two states can be retained for a long time (retention) up to many years. Incidentally it is worth noting that similar device structures, especially ECM, can be further engineered to achieve a threshold switching behavior suitable for selector application (more details in Chapter 5, RRAM-Based Coprocessors for Deep Learning) or even short-term retention to be exploited in spiking neural networks (SNNs, Chapter 17, Synaptic Realizations Based on Memristive Devices). An additional class of RRAM relies on uniform interfacial switching (Fig. 1.1A, bottom), in which the conductance scales with the junction area of the device, and the mechanism is related to a homogenous oxygen ion movement through the ...

Table of contents

• Cover image
• Title page
• Table of Contents
• Copyright
• List of contributors
• Preface
• Part I: Memristive devices for brain–inspired computing
• Part II: Computational memory
• Part III: Deep learning
• Part IV: Spiking neural networks
• Index