Performances Modulation Of TaO_x-based Memristor And Its Array

As the size of devices in chips approaches its physical limit,Moore’s law is coming to end.Meanwhile,the Von Neumann architecture,in which the information storage and computation are physically separated,limits the speed of the data storage and transmission,resulting in high energy consumption of the computing system.A memristor can be used as the basic device of the neuromorphic computing chip because of its computing-in-memory,low energy consumption and synaptic characteristics.In particular,memristors are compatible with complementary metal-oxide-semiconductor(CMOS)process and can be integrated to a large scale,which is expected to provide an effective solution to break through the Von Neumann bottleneck and improve the computing power of the chip.In this thesis,TaO_x,a typical transition metal oxide,is used as the material of resistive switching(RS)layer in the memristor.The performances modulation of the TaO_x-based memristor and its array are studied.The main research results are as follows:1.TaO_x-based memristors with the crossbar structure are fabricated by magnetron sputtering and photolithography.The device exhibits a self-rectifying RS effect with a rectifying ratio of 3×10~3.The rectifying effect is due to the Schottky barrier at the TaO_x-Pt interface,and the RS effect is due to the Schottky emission and the modulation of oxygen vacancies on the interface barrier.Based on the device performance,it is calculated that the array size can be 3574×3574,about 1.5 MB.2.After an electroforming process,the TaO_x-based memristor exhibits a nonvolatile RS effect with a retention of 1×10~4 s.The nonvolatile RS effect is accompanied by the generation of bubbles,which results from the growth and rupture of conductive filaments of oxygen vacancies.The analog RS type changes to the digital RS type by increasing the compliance current.The device can achieve an ON/OFF ratio of 10 and 35 distinguishable resistance states by pulse modulation.With non-overlapping pre-synaptic and post-synaptic pulses,the device successfully emulates the spiking-timing-dependent plasticity(STDP).3.By analyzing the mechanism of the weight updating in the TaO_x-based memristor,different pulse schemes are used to modulate the weight updating.To evaluate the impact of the weight updating on the classification accuracy and the energy consumption of the memristor,a three-layer neural network is used to classify handwritten digits(0-9).It is found that the linearity,range and symmetry of the weight updating have different impacts on the classification accuracy.Among these schemes,the classification accuracy reaches94%by using the depression pulse scheme,which is close to the ideal case.The energy consumption of a system based on the memristors is 438 times lower than that of a similar implementation using CMOS chip.4.TaO_x-based memristor arrays are designed and fabricated.The performances of the array,such as line resistance,electroforming voltage,resistance of high and low resistance states,SET voltage and RESET voltage,are analyzed.To reduce the electroforming damage of the array,the oxygen partial pressure of the magnetron sputtering and the thickness of the RS layer are modulated.5.TaO_x-based memristors are connected to a combined array after an electroforming process to avoid the reading interference between neighboring devices.The array is packaged by wire bonding and connected with the peripheral circuit to achieve the position recognition.