Designed Memristor Circuit for Self‐Limited Analog Switching and its Application to a Memristive Neural Network

Abstract

Self-limited switching is a technique that can control the memristor resistance up to a specific value by limiting excessive switching. [14,15] For example, in self-limited "set" switching, a series resistor (R S ) is connected to a memristor (M). The R S -M configuration is biased with a programming voltage (V P ) to switch the M from the HRS to the LRS. Before the switching happens, if the node voltage on M is V M , the V M is almost equal to the V P because the R HRS is much larger than the resistance of R S (V M ≈ V P ). As soon as the resistance state of M changes to the lower value (i.e., the set switching happens), the LRS is not uniform because of the stochastic characteristic of the memristor switching. Afterward, the V P is redistributed by the voltage divider effect, and the redistributed V M (V M ′) is lower than the V P . Then, the V M ′ leads to additional minor set switching of the M.Here, the amount of additional set switching depends on the V M ′; if the R LRS is relatively small, the V M ′ is also small, so there is less additional set switching. However, if the R LRS is relatively high, the V M ′ is also high and this leads to additional set switching. Therefore, if enough time is allowed, the set switching can spontaneously saturate to a certain value.Self-limited "reset" switching is complicated, but it is also possible to realize by adopting both series and parallel resistor components. [14,16] If such self-limited switching is applied to control the intermediate states, more uniform intermediate states can be achieved. Furthermore, if the resistance of the intermediate states is the same as the series resistor value, one can directly copy any resistance values from the R to the M under the self-limited switching regime. This suggests that the analog data programming will be more efficient and faster.In this study, we propose a novel analog data programming method that transfers reference analog resistance values to a target memristor directly via a stateful in-memory logic operation. We demonstrate the methodology by adopting an ideal memristor model design. The desired memristive behavior, characterized by a gradual set switching without a sudden current jump, was achieved using a Ti-doped NbO x charge trap memristor, but with some discrepancy in switching behavior compared to the theoretic ideal. [17] Afterward, we introduced parallel resistors in the programming circuit to reduce the discrepancy and bring the memristor behavior closer to that of the ideal model. Finally, a computational simulation was employed Analog memristors enable compact neuromorphic computing with low power consumption. One of the issues with the technology is slow precise analog data programming. In this study, a novel analog data programming method utilizing a self-limited set switching is proposed. The method can transfer any resistance values from reference resistors to the target memristor accurately inside a crossbar array by performing an appropriate voltage clocking. An ideal memristor model based on the me...

Implementation of Simple but Powerful Trilayer Oxide-Based Artificial Synapses with a Tailored Bio-Synapse-Like Structure

“…However, the development of energy-efficient neuromorphic hardware systems has been hindered by the limited performance of analogue synaptic devices . A specific memristor circuit was typically designed in order to achieve self-limited analogue switching, and thus its application to a memristive neural network . Analogue programming can provide the synapses with both analogue depression and potentiation by accurately controlling intermediate states.…”

“…33 A specific memristor circuit was typically designed in order to achieve self-limited analogue switching, and thus its application to a memristive neural network. 34 Analogue programming can provide the synapses with both analogue depression and potentiation by accurately controlling intermediate states. However, it is at the expense of an increased device fabrication complexity, and such accurate control is difficult because of the inherent stochastic characteristics of the resistance switching.…”

Discrete Memristor and Discrete Memristive Systems

“…In fact, researchers have used alternative techniques to design memristor devices based on mathematical models. Analog circuit implementation of memristors is important, and there are many different implementations [ 20 , 21 , 22 , 23 ]. Meanwhile, FPGA implementation has aroused much interest among researchers due to its easily programmable, reconfigurable, controllable, precise and better performance [ 24 , 25 ].…”

Defect-Engineered Electroforming-Free Analog HfOx Memristor and Its Application to the Neural Network

“…Another critical issue is the involvement of the random variation in the device characteristics between the cells and switching operations, which certainly impacts the performance of the network. , Such variations can be quantified by adopting a coefficient of variation (CV) that defines the ratio of the standard deviation (σ) to the mean value (μ). Song et al reported recently that CV of about ∼2% does not impact the system performance, demonstrating the defect-tolerant feature of the artificial neural network . In addition to these synaptic performances, conventional memory performance indicators, such as the retention and endurance of the memory cell, are still the crucial factors for the memristor to have in the neural network. − Most of the MIM-structured memristors are highly insulating at the as-fabricated state, which requires an electroforming (EF) step to operate them as a feasible memory or synapse.…”

“…reported recently that CV of about ∼2% does not impact the system performance, demonstrating the defect-tolerant feature of the artificial neural network. 22 In addition to these synaptic performances, conventional memory performance indicators, such as the retention and endurance of the memory cell, are still the crucial factors for the memristor to have in the neural network. 23−29 Most of the MIM-structured memristors are highly insulating at the as-fabricated state, which requires an electroforming (EF) step to operate them as a feasible memory or synapse.…”