Perpendicular Spin-Torque-Transfer Mechanism of Neuron: Integrate Vs. Stochastic Nature

Abstract

Achieving a higher integration density and a lower power consumption for neuromorphic devices, recently, spin neuron has been intensively researched. [ref . 1-3] In particular, 2-terminal perpendicular spin-transfer-torque (p-STT) neuron has been reported as a candidate of spin neuron. [ref 4-5] This neuron showed both stochastic nature as well as integrate nature, which is determined by reset voltage and spike amplitude (i.e. set voltage). However, the mechanisms between two different neuron nature has not been clearly proved.Here, two different mechanisms were reviewed by observing the dependency of neuron nature on reset voltage and spike amplitude. 2-terminal p-STT neuron, as shown in Fig. 1(a), was operated by two different reset voltage and spike amplitude (i.e. fully reset or partially reset) dislike p-STT MRAM, as shown in Fig. 1(b). Two different reset voltage and spike amplitude would have a different energy distribution from parallel state to anti-parallel state for integrate or stochastic nature; i.e. the energy level of parallel state for partially reset was higher than that for fully reset and the activation energy (i.e. parallel to anti-parallel state switching) for partially reset was lower than that for fully reset, as shown in Fig. 1(c).For fully reset case, the probability of switching from parallel to anti-parallel increased with the spike amplitude (i.e. negative bias magnitude) being able to produce a sigmoid function for stochastic neuron, as shown in Figs. 1(d)-(h). Otherwise, for partially reset case, the resistance margin between parallel and anti-parallel state was enhanced with the spike amplitude, being able to reduce a read error rate in fire circuit, as shown in Figs. 1(i)-(m). In addition, in our presentation, we will review in detail to different mechanism between integrate and stochastic nature by the grain size distribution of the MgO tunneling barrier in p-STT neurons. Particularly, it was found that a criteria between integrate and stochastic nature in p-STT neurons could be decided by the reset voltage amplitude; i.e. the ratio of stochastic (i.e. fully reset) to integrate (i.e. partially reset) decreased with increasing the reset voltage amplitude. ACKNOWLEDGEMENT This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. 2016M3A7B4910249)and the Brain Korea 21 PLUS Program in 2014. REFERENCE [1] Sengupta, Abhronil, and Kaushik Roy. "Spin-transfer torque magnetic neuron for low power neuromorphic computing." 2015 International Joint Conference on Neural Networks (IJCNN). IEEE, 2015.[2] Krzysteczko, Patryk, et al. "The Memristive Magnetic Tunnel Junction as a Nanoscopic Synapse‐Neuron System." Advanced Materials 24.6 (2012): 762-766.[3] Borders, William A., et al. "Analogue spin–orbit torque device for artificial-neural-network-based associative memory operation." Applied physics express 10.1 (2016): 013007.[4] Kondo, Kei, et al. "A two-terminal perpendicular spin-transfer torque based artificial neuron." Journal of Physics D: Applied Physics 51.50 (2018): 504002.[5] Kim, Dong Won, et al. "Double MgO-Based Perpendicular Magnetic Tunnel Junction for Artificial Neuron." Frontiers in Neuroscience 14 (2020). Figure 1