A Spin–Orbit‐Torque Memristive Device

Abstract

Memristors, demonstrated by solid-state devices with continuously tunable resistance, [1][2][3][4][5][6][7] have emerged as a new paradigm for self-adaptive networks that require synapse-like functions (artificial synapse, for example). Spin-based memristors offer advantages over other types of memristors because of their significant endurance and high energy efficiency. [8,9] Yet, it remains a challenge to build dense and functional spintronic memristors with structures and materials that are compatible with existing ferromagnetic devices. [10] Here, a memristive device based upon Ta/CoFeB/MgO heterostructures is demonstrated, which are commonly used in out-of-plane magnetized magnetic tunnel junctions. [11] To achieve the memristive function, a domain wall (DW) is driven back and forth in a continuous manner in the CoFeB layer by applying in-plane positive or negative current pulses along the Ta layer, utilizing the spin-orbit torque (SOT) that the current exerts on the CoFeB magnetization. [12][13][14][15][16][17] Hence, Memristors, demonstrated by solid-state devices with continuously tunable resistance, have emerged as a new paradigm for self-adaptive networks that require synapse-like functions (artificial synapse, for example). Spin-based memristors offer advantages over other types of memristors because of their significant endurance and high energy efficiency. Yet it remains a challenge to build dense and functional spintronic memristors with structures and materials that are compatible with existing ferromagnetic devices. Here, a memristive device based upon Ta/CoFeB/MgO heterostructures is demonstrated, which are commonly used in out-of-plane magnetized magnetic tunnel junctions (MTJ). To achieve the memristive function, a domain wall (DW) is driven back and forth in a continuous manner in the CoFeB layer by applying in-plane positive or negative current pulses along the Ta layer, utilizing the spin-orbit torque (SOT) that the current exerts on the CoFeB magnetization. Hence, the magnetization and consequently the anomalous Hall effect (AHE) resistance are modulated in an analog manner, being controlled by the pulsed current characteristics including amplitude, duration, and repetition number. The quasi-continuous AHE resistance variation is explained by the SOT-induced DW creep motion. These results pave the way for developing SOT-based energy-efficient neuromorphic systems.

Tailoring Multilevel‐Stable Remanence States in Exchange‐Biased System through Spin‐Orbit Torque

“…For the SOT induced magnetization switching in the HM/FM systems, domain nucleation and domain wall motion are the dominant process. [ 12–14 ] If the amplitude of current density J is less than J C (defined as the smallest current density required to completely switch the FM magnetization), there will be an incomplete switching of FM magnetization. [ 13,14 ] In HM/FM/AFM trilayers, if the partial FM magnetization is switched, the height of the positive (negative) biased loop will be changed since the AFM interfacial spins will also be switched to align with the FM magnetization.…”

“…[ 12–14 ] If the amplitude of current density J is less than J C (defined as the smallest current density required to completely switch the FM magnetization), there will be an incomplete switching of FM magnetization. [ 13,14 ] In HM/FM/AFM trilayers, if the partial FM magnetization is switched, the height of the positive (negative) biased loop will be changed since the AFM interfacial spins will also be switched to align with the FM magnetization. Thus, the change of the remanence and the heights of the sub‐loops in HM/FM/AFM trilayers could be delicately designed and realized by controlling the strength of SOT.…”

“…[ 11–13 ] It was found that the multilevel states with relatively large Hall resistance as readout signal could be achieved at J < J C through the incomplete switching of FM layer magnetization. [ 14,15 ] However, multilevel states could be more easily influenced by external field. If one combines the advantages of both AFM and FM systems, more significant and stable multilevel states in FM/AFM bilayers are naturally expected.…”

On-chip Learning In A Conventional Silicon MOSFET Based Analog Hardware Neural Network

“…Next we compare the energy and speed performance of on-chip learning on our proposed MOSFET synapse based FCNNN with that on FCNN designed using some other kinds of synaptic devices that have been proposed and implemented elsewhere -spin orbit torque driven domain wall based synapses (spintronic synapses) [7,49,50] and memristive oxide based RRAM synapses [51,52]. For fair comparison between the different synaptic devices, neural networks with the same architecture and number of nodes needs to be designed with different types of devices as synapses and they have to be trained on the same dataset using the same algorithm.…”

Spin-orbit torques and their associated effective fields from gigahertz to terahertz

“…which provide an efficient way to switch the magnetisation in systems containing heavy metals 6,7 or topological insulators [8][9][10] at room temperature. The manipulation of magnetic states by the current-induced torques provides novel capabilities for future technology -including memristors, building blocks for neuromorphic computing 11,12 .…”