Spintronic devices for neuromorphic computing
- PDF / 624,778 Bytes
- 3 Pages / 595.276 x 793.701 pts Page_size
- 48 Downloads / 244 Views
ly 2020 Vol. 63 No. 7: 277531 https://doi.org/10.1007/s11433-019-1499-3
Spintronic devices for neuromorphic computing 1
1
1
YaJun Zhang , Qi Zheng , XiaoRui Zhu , Zhe Yuan
1,2*
, and Ke Xia
1,2,3*
1
3
Center for Advanced Quantum Studies and Department of Physics, Beijing Normal University, Beijing 100875, China; 2 Peng Cheng Laboratory, Center for Quantum Computing, Shenzhen 518005, China; Shenzhen Institute for Quantum Science and Engineering and Department of Physics, Southern University of Science and Technology, Shenzhen 518055, China Received November 28, 2019; accepted December 17, 2019; published online February 17, 2020
Citation:
Y. J. Zhang, Q. Zheng, X. R. Zhu, Z. Yuan, and K. Xia, Spintronic devices for neuromorphic computing, Sci. China-Phys. Mech. Astron. 63, 277531 (2020), https://doi.org/10.1007/s11433-019-1499-3
In the past decades, significant progress has been achieved in artificial intelligence, which is now widely applied in image recognition, big data analysis, unmanned vehicle control and other cognitive tasks [1]. These applications nevertheless are highly energy consuming partly because of the mismatch between the neural-network-computing-based software implementation and the von Neumann architecture of present computers. One promising solution is developing neuromorphic chips without the so-called von Neumann bottleneck, which are suitable for performing the desired computation based on artificial neural networks (ANNs). So far there have been demonstrations of neuromorphic chips using the traditional semiconductor techniques achieving very low energy consumption [2]. Such chips usually require thousands of transistors to simulate one spiking neuron and employ separate synapses from neurons. Therefore, many attempts are made to develop neuromorphic chips with new materials and technologies, e.g. resistive memristors [3], phase change materials [4], and ferroelectricity [5]. Spintronic devices, which were extensively studied as memory units, intrinsically have the required dynamical properties of the basic elements—neurons and synapses—in an ANN [6]. The nonlinear dynamics of an artificial neuron can be replaced by magnetization dynamics since the latter is described by the nonlinear Landau-Lifshitz-Gilbert equation. *Corresponding authors (Zhe Yuan, email: [email protected]; Ke Xia, email: [email protected])
The tunable resistance used in magnetic memory devices is naturally artificial synapses, whose nonvolatility further reduces the power consumption. In addition, the remarkably 15 large endurance of magnetic devices (>10 ) [7] is particularly suitable for implementing reprogrammable neural networks. Then the focus of the past spintronics research, which aimed to develop memory devices, shall be significantly extended to achieve spintronic neural networks. For instance, as a memory device, one only needs to probe the static states before and after magnetic switching while the dynamical process of magnetization is essential in the neuromorphic computing as well as the stochastici
Data Loading...
