A sub-1-volt analog metal oxide memristive-based synaptic device with large conductance change for energy-efficient spike-based computing systems

Abstract

A memristor is a two-terminal 'memory resistor' electronic device, in which a metal oxide switching layer is sandwiched between two metal electrodes [1][2][3][4]. In general, memristors offer non-linear switching characteristics, and materials and process compatibility with advanced silicon manufacturing. These attributes have spurred the exploration of memristors as synaptic devices for realizing spike-based hardware learning systems that are capable of processing unstructured, temporal data [5][6][7][8][9][10]. However, for memristor-based technologies to be viable, the device should exhibit several key characteristics. It should have a compact nanoscale footprint, operate at a voltage close to 1V that is compatible with complementary metal oxide semiconductor (CMOS) technology, have reproducible electrical characteristics, and possess high switching speed to minimize the energy consumption [11]. Furthermore, the hardware integration of synaptic connections in advanced neural networks requires memristors with multiple resistive states [12,13]. These are challenging requirements and are difficult to implement without significant innovations.The phenomenological principle of memristor device operation is based on the change in the physical properties of a conductive filament (associated with the presence of oxygen vacancies) by applying an electric field across the metal oxide switching layer [14][15][16]. The resulting motion of the oxygen vacancies alters the device resistance between low (Set) and high (Reset) states, depending on the direction and the amplitude of the electric field. So far, a variety of structures from a large set of materials (various metal oxide switching layers and metal electrodes) have been studied in the literature [4,17,18]. Several key findings can be drawn from those studies regarding the performance, energy and scalability of this type of devices. The most important finding reveals the trade-off between the switching energy and the data retention time-that is often referred to as voltage-time dilemma [19]. This trade-off is associated with the energy barrier of the device structure. For example, devices made of metal oxides with small energy bandgap (E g ), such as titanium oxide (TiO x , E g~3 .4eV), generally exhibit low operating voltage and compromised data retention [20], while those with large bandgap, such as hafnium oxide (HfO x , E g~5 .4eV) demonstrate the opposite [21]. However, the fabrication of devices with bilayer switching stacks has shown to be effective in mitigating this trade-off. In particular, the improvement in data retention was obtained by the incorporation of an ultra-thin wide bandgap metal oxide capping layer (for example aluminum oxide) [22]. On the other hand, the addition of a reactive capping metal (for example titanium, hafnium, etc.) as an oxygen scavenging layer provided a pathway for reducing the operating voltage of the devices [23,24]. Despite significant advances, a sub-1V memristive device that simultaneously affords built-in analog behavior...

Memristive Synapses for Brain‐Inspired Computing

“…Memristive synapses with a similar working mechanism have been realized in IGZO x /IGZO y , HfO x /CeO x , TaO x /TaO y , TaO x /AlO x , TaO x /TiO 2 , TaO x /HfO 2 , TiO x /AlO x , MoO x /MoS 2 , and Al 2 O 3 /Nb x O y bilayer structures . It has been speculated that compared to synaptic devices composed of a single oxide layer, memristive synapses based on a bilayer structure possess more reliable performance due to controllable formation/rupture of conducting filament at the vicinity of the interface between these two layers (see Figure ) .…”

Multilayer Metal‐Oxide Memristive Device with Stabilized Resistive Switching

“…The traditional approaches to control the reproducibility of resistive switching include the formation of special electric field concentrators and appropriate selection of materials/interfaces in memristive device structure. In the latter case, bilayer or multilayer structures are formed, in which the switching oxide alternates with a barrier/buffer layer (layers) to control the migration of oxygen vacancies, with a layer of low dielectric constant to obtain nonlinear current–voltage ( I – V ) characteristics, or with a layer of higher/lower thermal conductivity for the removal/retention of heat in the switching area and to achieve analog switching character. To tune the resistive states with given accuracy, regardless of their native variation, adaptive programming of resistive state is actively employed by correcting the parameters of switching voltage pulses depending on the result of programming (so called write‐and‐verify approach).…”

Fast Spiking of a Mott VO2–Carbon Nanotube Composite Device

“…[3][4][5][6][7][8] In order to improve the energy-efficiency and speed of such systems it is desirable to control the pulse width and energy, and to produce the spiking using single scalable devices. 4,[9][10][11] For instance, adjusting the analog node weights of a neural network by small increments in order to enable high precision will require precise and tunable low energy pulses, especially in networks that use memristors such as phase change memory or oxide ionic resistive switches. 12,13 Partly owing to the absence of compact circuits that can produce such tunable low-energy pulses, even the best memristor-based neural networks have had to implement elaborate transistor-based circuits at every node of very large networks, making the system's efficiency far from ideal.…”