Probabilistic Computing Using Stochastic Magnetic Tunnel Junctions

Conventional electronics relies on deterministic operation of electronic devices, where stochastic behavior is attempted to be minimized. In contrast to this perspective, in 1981, R. P. Feynmann gave a suggestion of unconventional computing paradigm, so-called the probabilistic computing. This approach leverages intentionally enhanced probabilistic behavior of physical system within computing hardware to simulate physical phenomena that are inherently probabilistic. The demand for such probabilistic computers has risen recently as a rapid increase in computing tasks that can be efficiently addressed by probabilistic algorithms. Probabilistic bit (p-bit) is a fundamental unit constituting the probabilistic computer and recent studies have revealed that probabilistic spintronics devices, in particular, the stochastic magnetic tunnel junction (s-MTJ), shows promise for constructing the p-bit.<br/>In this talk, I show various proof-of-concepts for the spintronic probabilistic computers and also discuss the physics and engineering of s-MTJ. I first outline the basic properties and characteristics of the p-bit with s-MTJ and then showcase several demonstrations including combinatorial optimization [1], Boltzmann machine learning [2], quantum simulation [3], and Bayesian inference [4]. After that, I delve into the physics of the stochastic magnetic tunnel junction elucidating the time-domain [5,6] and time-averaged [7,8] properties. I also discuss advanced design of the s-MTJs [9-12] tailored for reliable, large-scale computers.<br/>These studies are carried out in collaboration with H. Ohno, S. Kanai, W. A. Borders, K. Hayakawa, K. Kobayashi, R. Ota, H. Kaneko, G. Finocchio, S. Datta, and K. Y. Camsari, and were partly supported by JST-CREST JPMJCR19K3, JST-AdCORP JPMJKB2305, JST-ASPIRE JPMJAP2322, MEXT X-NICS JPJ011438 and RIEC Cooperative Research Projects.<br/><br/><b>Reference</b><br/>[1] W. A. Borders et al., Nature 573, 390 (2019).<br/>[2] J. Kaiser et al., Phys. Rev. Appl. 17, 014016 (2022).<br/>[3] A. Grimardi et al., IEEE IEDM 2022, 22.4 (2022).<br/>[4] N. A. Singh et al., IEEE IEDM 2023, 12.1 (2023).<br/>[5] S. Kanai et al., Phys. Rev. B 103, 094423 (2021).<br/>[6] K. Hayakawa et al., Phys. Rev. Lett. 126, 117202 (2021).<br/>[7] K. Kobayashi et al., Appl. Phys. Lett. 119, 132406 (2021).<br/>[8] T. Funatsu et al., Nat. Commun., 13, 4079 (2022).<br/>[9] K. Y. Camsari et al., Phys. Rev. Appl. 15, 044049 (2021).<br/>[10] K. Kobayashi et al., Phys. Rev. Appl., 18, 054085 (2022).<br/>[11] K. Selcuk et al., Phys. Rev. Appl. 21, 054002 (2024).<br/>[12] R. Ota et al., arXiv:2405.20665 (2024).