Tri-state spintronic device, memory cell, memory array and read-write circuit

Abstract

The present disclosure provides a three-state spintronic device, a storage unit, an array and a read-write circuit. The three-state spintronic device includes from bottom to top: a bottom electrode, a magnetic tunnel junction and a top electrode; the magnetic tunnel junction includes : spin-orbit coupling layer, ferromagnetic free layer, barrier tunneling layer and ferromagnetic reference layer, three localized magnetic domain wall pinning centers and magnetic domain wall nucleation center; modulation of antisymmetric exchange, magnetic domain wall The pinning center is embedded at the interface between the heavy metal and the ferromagnetic free layer; the nucleation center of the magnetic domain wall is set at both ends of the ferromagnetic free layer; the current pulse flows through the spin-orbit coupling layer to generate spin current and inject into the ferromagnetic free layer. The effective field of the spin-orbit torque under the control of the electric field drives the displacement of the domain wall, and the displacement can be modulated by the number of pulses, the pulse width and the direction of the current, and has CMOS process compatibility and high reliability. The present disclosure also provides three-state read and write. The circuit and its ternary network computing application scheme realizes the high-performance GXNOR operation of ternary spintronic devices.

Description

technical field [0001] The present disclosure relates to the field of integrated circuits, and in particular, to a tri-state spintronic device, a storage unit, a storage array, and a read-write circuit. Background technique [0002] Neural network architecture has set off a research boom in academia and industry due to its unique advantages in image recognition, semantic recognition, and classification tasks. However, the training and recognition process of traditional convolutional neural networks requires a large number of floating-point and double-precision convolution operations. On the one hand, a large amount of data puts forward higher and higher requirements on the storage capacity of hardware. The multiplication and addition operation used in the product operation puts forward high requirements on the energy consumption and time of the operation, which makes the neural network training time for several days or even weeks. [0003] In order to optimize the problems ...

AI Technical Summary

Problems solved by technology

Compared with BNN, TNN has higher information capacity and does not increase the complexity of operation, so it has great application potential. There are few reports on the hardware based on ternary neural network operation.

Images