Ultra-low Power Domain Wall Device for Spin-based Neuromorphic Computing

Abstract

Artificial intelligence (AI) is gaining acceptance in a variety of applications. In order to develop an energy-efficient architecture for AI, neuromorphic computing, which draws inspiration from the human brain, is being investigated. This architecture is a complex interconnection of numerous synthetic neurons and synapses, which are the processing and memory units, respectively. To realize neuromorphic architecture, one needs to fabricate artificial neurons as well as synapses. Several charge and spin-based material systems are being studied towards developing neuromorphic architecture. However, magnetic domain wall (DW) devices are one of the most energy-efficient contenders1. Besides, utilizing spin-orbit torque to drive DWs provides more degrees of freedom in achieving higher energy efficiency. Therefore, in the present study, we are studying spin-orbit torque for driving the DWs in the proposed novel material engineering to further reduce energy consumption.In SOT driven DW motion, the power can further be reduced either by increasing spin Hall angle or by reducing pinning field while keeping spin Hall angle at optimized values. Since W in β phase results in the highest spin Hall angle (θSH)2, we decided to study β-W-based spin Hall layer. However, β-W is obtained by depositing the W films at low power and high Ar gas pressure (HP). These depositions conditions result in granular W films, which in turn results in (a) more defects, (b) higher critical current, (c) larger power, and (d) lower TMR2,3. To resolve these, we have proposed the concept of the dual W layer, where a low-pressure (LP) W, which is expected to be smoother (lesser grains), is deposited between β-W and ferromagnetic layers. The studied stack involves dual W spin Hall layer (SHL) (6 nm)/ CoFeB (1 nm)/ MgO (~1 nm)/ Ru (2 nm). The studied stack is schematically illustrated in figure 1 (a).Firstly, we characterized the thin film samples and observed that the HP3LP3 (this means SOT layer consists of high pressure deposited W (3 nm)/ low pressure deposited W (3 nm)) dual-layer sample exhibits the best structural, magnetic, and electrical characteristics. Therefore, we fabricated DW devices from the HP3P3 sample. Moreover, we prepared DW devices from single LP6 and HP6 spin Hall layers for comparison purposes. The details of the measurements at the thin film level will be presented in detail during the conference.Subsequently, we studied the DW dynamics in our samples using Kerr microscopy. First, we saturated the DW devices with a large magnetic field. Then, we applied a reversed magnetic field to insert DW in the DW devices. Once the DW is inserted into the devices, we applied current pulses of various amplitudes and observed the DW motion. When current pulses were applied the DW moves in a direction opposite to the current (please refer to figure 1 (b-d). Moreover, as the amplitude of the current decreases the DW velocity decreases. Most importantly, the DW moves at the current density as small as 107 A/m2, as shown in figure 1 (d). The detailed measurements revealed that the DW can be moved to a minimum current density of 106 A/m2. Note, a small OOP magnetic field of -1 mT, which is approximately one-third of the coercivity of the devices, was applied during these measurements. On contrary, we observed that devices with a single spin Hall layer do not show the DW motion.Further harmonic Hall measurements and pinning field measurements showed that ultra-low pinning field (figure 2) in addition to the optimum value of spin Hall angle results in DW motion at such low currents in our dual-layer devices. We have also compared the energy consumption in our devices with respect to the current state-of-art and found a significant reduction in energy consumption in our devices4-6.In addition, we have proposed the design of DW-based neuron and synaptic devices. Moreover, we have proposed and studied the synaptic functions in meander wire, in which two neighboring segments meet at a small offset of “d”. We also demonstrated the working stochastic, which is for d= 15%, and non-stochastic, which is for d=50%, type synapses in our dual-layer meander devices. In conclusion, the observation of such energy-efficient DW motion is a significant contribution to the field of neuromorphic computing. The results in detail will be presented during the conference. **