Study On Spin-orbit Torque Devices For In-memory Analogue Computing

With the advent of Internet of Things(Io T)and artificial intelligence(AI),digital computing based on von Neumann architecture consumes a lot of energy and time when processing amounts of information.On the other hand,Moore’s Law is gradually coming to an end,and the performance of silicon-based transistors cannot be improved greatly.Therefore,in-memory analogue computing(IMAC)based on emerging non-volatile devices is gradually returning as a new paradigm,which is expected to achieve energy-efficient and time-efficient computing.Among them,spin-orbit torque(SOT)devices have received extensive attention due to their separation of reading and writing paths,low power consumption,high speed,high stability,and high durability.Currently,analogue arithmetic operations and neuromorphic computing based on analogue artificial synapses are two important research directions in IMAC.The analogue arithmetic operations are mainly realized by circuits composed of multiple separate units,which has the problems of high area overhead and high time delay.Analogue artificial synapses are mainly based on emerging non-volatile devices to achieve product operations.However,the linearity and symmetry of conductance modulation is generally hard to achieve for different kinds of artifcial synapse devices and there are also problems such as the requirement of initialization or external magnetic field assistance,which affect the accuracy of artificial neural networks and limit their applications.Aiming at the above problems,firstly,modulating the magnetization states change of a magnet by SOT was studied,and a linear relationship between the magnetic field and the resistance of SOT device(RH)was established,thus realizing the linear sensing of the magnetic field via RH or the linear modulation of RH by the magnetic field.Then based on the linear sensing of the magnetic field,the sensing of three-dimensional magnetic fields was realized.Moreover,the current-induced magnetic field sensing was studied and a linear relationship between the currents and RH was established,thus realizing the linear sensing of the currents or the linear modulation of RH by the currents.Then,based on the linear sensing of currents,in-memory analogue multipliers and adders/subtractors were made,and the amplitude modulation of signals as well as the edge extraction of pictures was realized.Finally,the artificial synapses were realized based on the linear modulation of the RH by currents.First,based on the magnetic domain wall motion(nucleation)in Ta/CoFeB/MgO/Ta(W/CoFeB/MgO/Ta),change of magnetization states of a magnet via SOT was investigated.A linear relationship between the magnetic field and RH was established.Then,based on the linear sensing of the magnetic field,the sensing of three-dimensional magnetic fields was realized by using a single SOT device.Moreover,based on the study of current-induced magnetic fields sensing,a linear relationship between currents and RH was established.Further,based on the linear sensing of currents,a four-quadrant in-memory analogue multiplier and an in-memory analogue adder/subtractor were implemented.The analogue multiplier could obtain the products of two input current signals by measuring the anomalous Hall voltage(UH)of the SOT device,and the amplitude modulation was realized based on the analogue multiplier.The analogue adder/subtractor utilized Kirchhoff’s current law to realize multi-inputs analogue addition/subtraction,and the edge extraction of the picture was realized.Finally,based on the modulation of RH by currents with high linearity and symmetry,the devices were used as synapses,and the resistances were used as weights to construct an artificial neural network,and the recognition rate of handwritten digits was as high as 95%.The IMAC system proposed in this thesis reveals the feasibility of in-memory analogue computing based on the sensing fuction of spintronic devices,and opens up a new way for spin technology to be used in new computing paradigms.