From MRAM to SP-MTJ: An Ultrathin Mo Insertion in Magnetic Tunnel
Junction-based P-bits for Solving Combinatorial Optimization Problem
Yu-Hsuan Lin1#, Ching Shih1,2#, Ming-Chun Hong1,2, Chen-Yu Yang1,2, Guan-Long Chen2, Hsin-Han Lee2,
Yu-Chen Hsin2, Chiao-Yun Lo2, Sin-You Huang2, Ting-Syun Huang1, Cheng-Yi Shih2, Shan-Yi Yang2,
I-Jung Wang2, Yao-Jen Chang2, Shih-Ching Chiu2, Yi-Hui Su2, Chih-Yao Wang2, Kuan-Ming Chen2,
Ho-Lin Tsai2, Jeng-Hua Wei2, Shyh-Shyuan Sheu2, Wei-Chung Lo2,
Shih-Chieh Chang2, and Tuo-Hung Hou1,3*
1
Institute of Electronics, National Yang Ming Chiao Tung University, Hsinchu, Taiwan
2
Electronic and Optoelectronic System Research Laboratories, Industrial Technology Research Institute, Hsinchu, Taiwan
3
Industrial Academia Innovation School, National Yang Ming Chiao Tung University, Hsinchu, Taiwan
#
Both authors are contributed equally. *E-mail:
[email protected]2025 International VLSI Symposium on Technology, Systems and Applications (VLSI TSA) | 979-8-3315-4312-9/25/$31.00 ©2025 IEEE | DOI: 10.1109/VLSITSA64674.2025.11046915
Abstract—The superparamagnetic tunnel junction (SP-MTJ) HC and switching voltage degrade significantly for the Mo-
is a promising candidate for p-bit-based stochastic computing, insertion-based MTJ device compared to the conventional
offering significant area efficiency over traditional CMOS STT-MRAM. Furthermore, as the thickness of Mo increases,
approaches. This study presents a simple method to enable SP- the Δ (Fig. 5a) and the HC (Fig. 5b) decrease significantly in
MTJ functionality by adding a molybdenum (Mo) layer above statistical analysis, indicating the degradation of the
the free layer in conventional spin-transfer-torque MRAM ferromagnetic property due to limited boron diffusion [5]. The
(STT-MRAM). The thermal stability factor (Δ), modulated by switching phase diagram (SPD) measurement is utilized for
Mo layer thickness, is analyzed to guide SP-MTJ design for
extracting the main switching factors (Fig. 6a) [6].
stochastic oscillatory behavior. This architecture is ideal for
Quantitatively, all switching factors decrease with increasing
solving combinatorial optimization problems (COPs), such as
the travelling salesman problem (TSP).
Mo insertion thickness (Fig. 6b). As the Mo thickness exceeds
1.5 Å, the Poisson probabilistic curve (Fig. 7a) with stochastic
I. INTRODUCTION oscillation in the SP-MTJ at Hext = -34 Oe shows a duty cycle
(probability of the P state) of approximately 50% (Fig. 7b) and
The p-bit circuit, leveraging SP-MTJ devices with low a mean switching time of 110 ms (Fig. 8). We extend SP-MTJ
energy barriers, is driven by thermal fluctuation [1]. It is used behavior to solve the 7-city TSP using stochastic discrete
in technologies like random number generators for spiking Hopfield neural network (DHNN) algorithm [7] (Fig. 9).
neural networks (SNNs) and stochastic computing for solving Compared to the conventional DHNN (Fig. 10a), the
COPs (Fig. 1) [2], where random oscillations are crucial. stochastic DHNN (Fig. 10b) achieves a high success rate
Compared to conventional CMOS oscillation circuits, the SP- (>50%) in finding the optimal solution. Fig. 11 presents the
MTJ cell is more area- and energy-efficient [3], with routing schedule for the 7-city TSP with an optimal distance
oscillations driven by the device intrinsic properties. of 283.68, consistent with the solution from the Concorde
Traditional perpendicular magnetic anisotropy (PMA)-based solver [8].
SP-MTJs are typically achieved by increasing the free layer
thickness [4]. However, this increases saturation IV. CONCLUSION
magnetization (MS) and reduces Δ simultaneously, affecting In this study, we demonstrated the SP-MTJ with a thin Mo
stochastic oscillation. In this study, we propose a novel layer above the free layer to weaken the PMA, reducing Δ and
approach: inserting Mo on the top of the free layer to weaken HC. This enables low-voltage, low-current operation in a
PMA in conventional STT-MRAM. The device achieves low single MTJ cell, resulting in energy- and area-efficient p-bit
Δ without increasing MS, enabling random oscillations at low devices. Our simulations show the strong potential of the
voltage and meeting SP-MTJ application requirements. This proposed device for solving COPs in real-world applications.
stochastic behavior improves success rates in solving COPs.
REFERENCES
II. DEVICE DESIGN OF SP-MTJ [1] M. Bapna et al., "Current control of time-averaged magnetization in
Fig. 2a shows the transmission electron microscopy (TEM) superparamagnetic tunnel junctions," Appl. Phys. Lett., vol. 111, no. 24, p.
243107, 2017.
image and energy-dispersive X-ray spectroscopy (EDX)
[2] B. Parks et al., "Superparamagnetic perpendicular magnetic tunnel junctions for
mapping of conventional STT-MRAM with Ta-based true random number generators," AIP Adv., vol. 8, no. 5, p. 055903, 2018.
composite free layer (Ta-CFL). Previous studies have [3] M.-H. Wu et al., "Compact probabilistic Poisson neuron based on back-hopping
mentioned that the different metal insertions would potentially oscillation in STT-MRAM for all-spin deep spiking neural network." 2020 IEEE
affect the PMA of the free layer [5]. Hence, in our subsequent Symp. on VLSI Tech., JFS4.2, 2020.
experiments, we vary the thickness of Mo above the Ta-CFL [4] W. Borders et al., "Integer factorization using stochastic magnetic tunnel
(Fig. 2b). This significantly reduces the energy barrier junctions," Nature, vol. 573, pp. 390-393, 2019.
between the antiparallel (AP) and parallel (P) states, enabling [5] M. Li et al., "Influence of inserted Mo layer on the thermal stability of
perpendicularly magnetized Ta/Mo/Co20Fe60B20/MgO/Ta films," AIP Adv., vol.
easier transitions in the MTJ. Increased Mo thickness 6, p. 045107, 2016.
maintains the TMR and low RA (Fig. 3a-b), while The M-H [6] G. Mihajlović et al., "Origin of the resistance-area-product dependence of spin-
loop shows slight degradation in MS and coercive field (HC), transfer-torque switching in perpendicular magnetic random-access memory
indicating the superparamagnetic properties (Fig. 3c). cells," Phys. Rev. Appl., vol. 13, no. 2, p. 024004, 2020.
[7] M.-C. Hong et al., "In-memory annealing unit (imau): Energy-efficient (2000
III. CHARACTERISTICS OF SP-MTJ tops/w) combinatorial optimizer for solving travelling salesman problem," 2020
IEEE International Electron Devices Meeting (IEDM), pp. 21.3.1-21.3.4., 2021.
The basic R-H (Fig. 4a) and R-V (Fig. 4b) curves of the [8] D. L. Applegate et al., "The traveling salesman problem: A computational study,''
SP-MTJ device with Mo thickness varying from 0 to 1 Å. The Princeton university press, 2006.
979-8-3315-4312-9/25/$31.00
Authorized licensed use limited to: ©2025
NationalIEEE
Yang Ming Chiao Tung University. Downloaded on July 25,2025 at 12:30:15 UTC from IEEE Xplore. Restrictions apply.
(a) (b)
Fig. 1 Schematic of the SP-MTJ-based P-bit circuit and its hardware Fig. 2 Material analysis of conventional STT-MRAM and optimization
applications. The intrinsic stochastic behavior of the SP-MTJ, induced by strategy of SP-MTJ. (a) TEM image and EDX mapping of conventional MTJ
environmental thermal noise, enables its use in true random number with Ta-CFL, and (b) energy diagram differences between MTJs with and without
generators, spiking neural networks, and solving combinatorial Mo insertion. Mo insertion reduces the energy barrier to EB below 15kBT,
optimization problems. enabling oscillatory behavior.
(a) (c) (a) (b)
(b)
Fig. 3 Analysis of blanket film samples. The (a) TMR ratio, (b) RA values, Fig. 4 Characteristics of MTJ devices. (a) Normalized R-H loops and (b)
and (c) M-H loops were measured on a blanket film with varying Mo R-V characteristics with a typical house-shaped curve were measured for
thicknesses inserted above the CFL. As the Mo thickness increases, TMR MTJ devices with varying Mo insertion thicknesses. Both the HC and the
shows no degradation, while HC significant decreases, indicating a reduction switching voltage decrease as the Mo thickness increases, indicating a
in magnetic anisotropy caused by Mo insertion. reduction in the device's magnetic stability.
(a) (a) (b) (a)
(b) (b)
Fig. 5 Statistical data of MTJ Fig. 6 SPD diagram and extraction of switching factors. Comparison of Fig. 7 Basic characteristics of SP-
devices. (a) HC and (b) Δ as functions (a) the SPD diagram and (b) the three major switching factors of the MTJ MTJ with 1.5 Å Mo insertion. (a)
of Mo thickness. Δ decreases with the device with varying Mo thickness. Both the switching field (HSW) and Poisson probabilistic curve and (b)
reduction in of HC, indicating a loss in statistical switching factors significantly decrease, suggesting a weakening oscillation behavior are observed with
deterministic switching performance. of the device’s magnetic properties and reduced switching efficiency. 1.5 Å Mo insertion.
(a)
(b)
Fig. 8 Mean switching time analysis Fig. 9 Schematic of the algorithm used Fig. 10 Statistical solution for Fig. 11 The route schedule for the
of the SP-MTJ sample. The mean to solve the TSP. Illustration of the HNN each trial. Distribution of solution 7-city TSP. Illustrations of the
switching time analysis of the SP-MTJ with deterministic and stochastic paths by (a) DHNN and (b) optimal travel routes for the 7-city
with a 1.5 Å Mo insertion layer shows processes. The P-bit exhibits its potential stochastic DHNN, with the latter TSP, with each point representing a
110 ms with Δ ≈ 14. for stochastic DHNN. achieving a success rate > 50%. city location.
Authorized licensed use limited to: National Yang Ming Chiao Tung University. Downloaded on July 25,2025 at 12:30:15 UTC from IEEE Xplore. Restrictions apply.
