Source: The content is reprinted from the official account of Semiconductor Industry Observer (ID: icbank) Yuezhi.com
,
author:
Li Huanglong, Yang Yifei
, etc. Thank you.
People's continuous exploration of the working mechanism of the human brain has given rise to the research of efficient brain-like computing. With the update and iteration of semiconductor technology, neuromorphic devices have brought researchers the dawn of efficient brain-like computing. Opportunities and challenges coexist, and there will be many difficulties in the working mechanism and materials of neuromorphic devices. The Brain-like Computing Team of the Department of Precision Instruments of Tsinghua University has conducted in-depth research on this and achieved a series of breakthroughs.
From a functional point of view, the human brain is a powerful "information processing system", but it is like a "black box". To this day, people still know very little about its mechanism. Modern neuroscience shows that the human brain contains about 86 billion neurons, which is roughly the same as the number of stars in the Milky Way, and the number of synapses connecting neurons is 1,000 times the number of neurons. This shows the high complexity of the brain structure, which has largely delayed people's pace of exploring the working mechanism of the brain. Fortunately, the rapid development of neuroscience in recent decades is bringing more inspiration to deciphering the "black box".
While trying to understand the brain, people are also trying to build brain-like computing systems that function similarly to the brain. Since the 20th century, semiconductor technology and computer science have made great progress, which has set off a wave of building electronic brains. At present, computers based on the von Neumann architecture are still the mainstream and can replace humans to perform fixed rule-based tasks faster and more accurately. With the increasingly urgent demand for more advanced artificial intelligence, people have gradually realized that von Neumann architecture computers still lag behind the human brain in adaptability, fault tolerance and generalization capabilities, and people urgently need to find a more brain-like computing paradigm. Neuromorphic electronics has brought hope to achieve this goal. It aims to realize computing functions by simulating neural behavior, and simulating neural behavior requires semiconductor devices with similar characteristics to its realization, that is, neuromorphic devices.
━━
━━
Memristors and neuromorphic devices
Back in 1971, Leon O. Chua predicted in his article “Memristor-The Missing Circuit Element” that in addition to the three basic circuit elements of resistance, capacitance and inductance, there is another basic element - memristor, as shown in Figure 1. In 2008, 37 years after the prediction, HP Labs published the article “The Missing Memristor Found” in Nature, announcing that the memristor phenomenon was observed in a new type of micro-nano semiconductor device. Therefore, Leon O. Chua is also known as the “Father of Memristor”.
The resistance adjustability and memory characteristics of memristors (the current resistance value depends on the voltage/current excitation history) are very similar to the long-term plasticity of synapses, and synaptic plasticity is the basis of the brain's learning and memory functions. Since then, memristor devices and neuromorphic devices have almost become two interchangeable concepts, and neuromorphic electronics has entered a stage of rapid development. In addition to biological similarities, memristors (circuits) have potential advantages over traditional complementary metal oxide semiconductor (CMOS) devices (circuits) in terms of miniaturization, storage density, and power consumption.
In the past decade, neuromorphic devices based on the memristor effect have made many important progress. In terms of material technology, from inorganic to organic, from ordinary materials to quantum materials, from ferroelectric materials to ferromagnetic materials, from bulk materials to low-dimensional materials, etc., all show their own unique neuromorphic characteristics. In terms of function, memristors can simulate more and more synaptic plasticity functions, and are no longer limited to synaptic simulation, but can also simulate neuronal functions, which creates the possibility for the realization of all-memristor neuromorphic circuits. In addition to single device research, research on neuromorphic integrated circuits that mix traditional transistors and memristors has also made gratifying progress in terms of scale and function.
At present, neuromorphic technology based on memristors still faces many challenges. In terms of bionic capabilities, neuromorphic devices are not yet able to simulate biological neural tissues to a high enough degree. Most of them can only simulate macroscopic electrophysiological behaviors at the cellular scale, but cannot simulate mechanisms at the subcellular scale. In terms of computing functions, the functions of neuromorphic circuits are still relatively rudimentary. They are mostly used as hardware accelerators for artificial deep neural network algorithms, and cannot fully demonstrate the unique temporal dynamics computing advantages of biological neural networks. In terms of integration technology, the performance of memristors cannot fully meet the needs of large-scale integration. Generally, they still require the assistance of traditional transistors, and cannot fully reflect the miniaturization and low power consumption advantages of memristors.
Based on the above challenges, the Brain-Inspired Computing Team of the Precision Instrument Department of Tsinghua University (hereinafter referred to as the Team) conducted in-depth research on the working mechanism and materials of neuromorphic devices, and proposed original and groundbreaking solutions.
━━
━━
Three-port memristor: synaptic long-term and short-term
Plasticity coexists in a single device
By drawing on the principles of neurobiology, the team realized that even in simulating the most basic synaptic plasticity function, neuromorphic devices have the problem of insufficient bionics, which leads to limited device functions. Specifically, in addition to long-term plasticity, synapses also have short-term plasticity, and the two occur in completely different regions and mechanisms at the subcellular scale, as shown in Figure 2 (a). The coexistence of the two is also one of the important sources of the rich temporal dynamics of biological neural networks. However, in traditional neuromorphic devices, the long-term plasticity and short-term plasticity of synapses are simulated by homologous physical mechanisms and occur in the same area of the device, resulting in the two plasticities being mutually exclusive and unable to coexist.
To address this problem, the team proposed a three-terminal memristor solution, as shown in Figure 2 (b). The three-terminal memristor has two natural spatial regions, namely the source-drain channel (the channel between the source and drain) and the gate capacitance; it also naturally has two physical effects, namely the short-term gate capacitance effect and the long-term gate memristor effect. The gate capacitance effect directly acts on the channel, changing the channel conductivity in a short time; while the gate memristor effect comes from the change in the state of the gate dielectric, which indirectly affects the channel conductivity. In this way, the long-term and short-term plasticity of the synapse can be simulated by different physical mechanisms in different regions, thereby achieving coexistence in a single device, as shown in Figure 2 (c). This study provides a universal basic component for brain-like computing characterized by dynamic time complexity. The relevant research results were published in the internationally renowned journal Advanced Materials under the title "Truly Concomitant and Independently Expressed Short- and Long-Term Plasticity in a Bi2O2Se-Based Three-Terminal Memristor".
━━
━━
Based on tellurium conductive channel generation and fracture resistance switching
Mechanism-based all-memristor device integration solution
In terms of device integration, large-scale neuromorphic synaptic networks still generally have the problem of crosstalk between devices, which has a serious impact on the reliability of neural network reasoning and learning processes. The traditional solution is to connect a transistor switch in series with each memristor to open the channel in the working state and block the interference from the non-working device. It is worth noting that the transistor is a device with a large area and high power consumption, and the intervention of the transistor switch will inevitably make the advantages of memristor technology in miniaturization and low power consumption unable to be reflected.
To address this problem, the team proposed a scheme for integrating all-memristive devices based on the generation and fracture resistance switching mechanism of the conductive channel of the single-element semiconductor tellurium (Te). Single-element tellurium semiconductor materials have the characteristics of electrochemical activity, low melting point, low thermal conductivity and low electrical conductivity. They can be used to construct solid-state electrochemical devices that can achieve transient switching under large driving currents. Their functions are similar to transistors, but they have simpler structures and stronger miniaturization capabilities, as shown in Figure 3 (a). The mechanism for achieving this function is that the tellurium conductive channel is first generated under the action of electrochemistry and then melted under the action of large current Joule heat. At the same time, the same device can also achieve stable regulation of the resistance state under small currents, simulating the long-term plasticity of synapses, as shown in Figure 3 (b). This is related to the lower conductivity of tellurium semiconductors than metals. High current switching and low current synaptic functions are the basic requirements for devices in neuromorphic synaptic networks, but they are a basic contradiction in traditional memristors. Generally speaking, large currents will make the material structure change more thoroughly, and the resistance state will be maintained for a longer time accordingly, and vice versa. This is the "current-resistance holding time" contradiction encountered by memristors in synaptic network applications. The team was able to resolve it by cleverly utilizing the electrical, electrochemical and thermal properties of tellurium semiconductors. The team also further demonstrated the complete "switch/synapse" series device function using a pair of identical devices, as shown in Figure 3 (c), providing a valuable reference for homogeneously integrated memristor neuromorphic synaptic networks. The relevant research results were published in the internationally renowned journal Nature Communications under the title "A new opportunity for the emerging tellurium semiconductor: making resistive switching devices".
━━
━━
latest progress
In terms of neuromorphic device research, based on the above-mentioned interim results, the team is gradually conducting research on neuromorphic device integrated circuits, in order to simulate the computing function of the brain at the system level and achieve cross-scale collaborative optimization of materials-devices-systems. In addition, the team was invited to publish a review article entitled "A Marr's Three-Level Analytical Framework for Neuromorphic Electronic Systems" in the internationally renowned journal Advanced Intelligent Systems, which summarizes its own insights on the development of brain-like computing systems and research on progress in this field. At present, the team has developed a 10k scale integrated array based on analog memristors, low operating current (off-state current as low as 10-12A, on-state current as low as 10-8A) and crosstalk resistance, as shown in Figure 4, and is deploying a series of new brain-like computing algorithms on this platform.
Acknowledgements: We would like to thank the National Natural Science Foundation of China (Project No. 61974082, 61704096, 61836004), China Association for Science and Technology Young Talent Support Project (Project No. 2019QNRC001), Beijing Center for Brain Science and Brain-Inspired Research, Tsinghua University-IDG/McGovern Institute for Brain Research "Brain+X" Project, National Key R&D Program of China (Project No. 2018YFE0200200), Beijing Science and Technology Plan Project (Project No. Z181100001518006, Z191100007519009), Suzhou-Tsinghua Innovation Leading Action Special Project (Project No. 2016SZ0102), and Tsinghua University-CETC Joint Research Center for Brain-Inspired Computing for support.
This article was published in the November 2021 issue of IEEE Spectrum Chinese version, "Technology Overview".
Li Huanglong
:
Associate Professor, Department of Precision Instruments, Tsinghua University/Center for Brain-Inspired Computing, Tsinghua University/Beijing Center for Brain Science and Brain-Inspired Research
.
Yang Yifei:
PhD student at the Department of Precision Instruments, Tsinghua University/Center for Brain-Inspired Computing, Tsinghua University
.
Guo Yunpeng:
PhD student at the Department of Precision Instruments, Tsinghua University/Center for Brain-Inspired Computing, Tsinghua University
.
Wang Xinxin:
PhD student at the Department of Precision Instruments, Tsinghua University/Center for Brain-Inspired Computing, Tsinghua University
.
Kong Ruikai:
PhD student at the Department of Precision Instruments, Tsinghua University/Center for Brain-Inspired Computing, Tsinghua University
.
He Yuhan:
PhD student at the Department of Precision Instruments, Tsinghua University/Center for Brain-Inspired Computing, Tsinghua University
.
Miao Chenfei:
Undergraduate student in the Department of Electronic Engineering, Tsinghua University
.
*Disclaimer: This article is originally written by the author. The content of the article is the author's personal opinion. Semiconductor Industry Observer reprints it only to convey a different point of view. It does not mean that Semiconductor Industry Observer agrees or supports this point of view. If you have any objections, please contact Semiconductor Industry Observer.
Today is the 2928th content shared by "Semiconductor Industry Observer" for you, welcome to follow.
Semiconductor Industry Observation
"
The first vertical media in semiconductor industry
"
Real-time professional original depth
Scan the QR code
, reply to the keywords below, and read more
Wafers｜ICs｜Equipment
｜Automotive Chips｜Storage｜TSMC｜AI｜Packaging
Reply
Submit your article
and read "How to become a member of "Semiconductor Industry Observer""
Reply
Search
and you can easily find other articles that interest you!
京公网安备 11010802033920号