Introduction

The spintronics by spin torque has emerged as a promising technology for the next generation memory and logic due to its non-volatility, high-speed operation, and superior endurance1,2,3,4. However, its ultra-fast switching capabilities are facing challenges due to the exponential increase in write current density as pulse widths approach the sub-nanosecond regime5,6,7. High current demands not only escalate energy consumption but also necessitate large-footprint transistors for current regulation8,9,10,11, thereby compromising storage density. Furthermore, high-current operation degrades endurance and reliability12,13,14,15. Thus, state-of-the-art spintronics, both spin-transfer torque magnetic random access memory (STT-MRAM)16 and spin-orbit torque MRAM (SOT-MRAM)17 variants—demonstrate practical operation timescales of 10–20 ns at the circuit level. Consequently, achieving sub-nanosecond switching speeds remains a pivotal challenge for realizing high-speed spin torque device with high efficiency18,19,20.

Parallel writing unlocks orders-of-magnitude speed acceleration in microelectronics by enabling simultaneous programming of device arrays. This strategy has been successfully demonstrated in NAND flash technology, where simultaneous programming of ~105 cells enables effective write speeds of 3 × 108 bit/s21,22—three orders of magnitude faster than single-cell operation (2 × 105 bit/s)23. Recent efforts to adapt this paradigm to spin torque devices have explored multi-cell SOT-MTJ arrays on shared spin source channels24,25,26,27,28. The STT-assisted SOT24,25,26 and voltage-controlled magnetic anisotropy (VCMA) modulation27,28 are used to enable selective cell addressing. However, these implementations remain fundamentally limited to write one MTJ. The parallel writing of multiple MTJs is still challenging.

The difficulty of the parallel writing for MTJs is the switching requirement of bipolar current in conventional MTJ. In this work, we propose a model to realize unipolar MTJ by controlling the interlayer exchange coupling via unipolar current in an orange peel coupling system. A transient Joule heating from unipolar current pulses raises the temperature, induces interfacial ferromagnetic (FM) and antiferromagnetic (AFM) coupling transitions and decay of stray field in an orange peel structure (Fig. 1a), which determines the magnetic switching direction (Fig. 1b). VCMA effect (VG) modulates the critical switching current density while preserving unipolar operation (Fig. 1c), which can enable independent state programming in parallel writing architecture (Fig. 1d). Our work provides an alternate framework and pathway for achieving high-speed, low energy and high-capacity spintronics devices and system.

Fig. 1: Parallel writing of MTJs.
Fig. 1: Parallel writing of MTJs.
Full size image

a Schematics of thermally modulated AFM/ FM coupling in an orange peel coupled system. Ferromagnetic coupling dominates at low temperature T1, while antiferromagnetic coupling prevails at higher temperature T2. b Unipolar current-driven magnetization switching. A single-polarity write current density (Jwrite) controls the MTJ’s resistance state by thermally inducing AFM/FM coupling transitions. c VCMA modulation. Gate voltage (VG) tunes the critical Jwrite threshold, enabling VG -selectable high/low resistance states. d Parallel programming architecture. Orthogonal control of Jwrite and VG (state selection) enables simultaneous multi-bit writing in unipolar MTJ arrays.

Results

Model of tuning interlayer exchange couplings by unipolar current

We propose a unipolar-writable MTJ model based on current-amplitude-modulated AFM/FM interlayer coupling transitions. The switching mechanism originates from interlayer exchange coupling control in the free layer (FL)/MgO/reference layer (RL) structure, where the current amplitude mediates AFM-FM coupling transitions through controlled Joule heating. The final magnetic state is determined by the joint effects of interlayer coupling strength and the stray field from the synthetic antiferromagnet (SAF) during the write pulse relaxation phase.

The interfacial coupling is modeled using an orange-peel coupling model with periodic roughness29, as schematized in Fig. 2a. The total system energy (E) comprises three key contributions: interlayer exchange energy (Eex), interfacial anisotropy energy (Ean), and magnetostatic energy (Emag). The coupling type is governed by the energy difference ΔE = E ↑ ↑ - E ↑ ↓, where negative (positive) ΔE indicates FM (AFM) coupling. (Complete mathematical formulation using single and doth sinusoids and how it is deduced are provided in Supplementary Note S1.)

Fig. 2: Mechanisms and experimental validation of magnetic coupling transition.
Fig. 2: Mechanisms and experimental validation of magnetic coupling transition.
Full size image

a Schematics of the magnetostatic interaction in an orange peel couplings system. b The energy difference (ΔE) between AFM coupling and FM coupling vs. magnetic anisotropy constant Ku. c Temperature dependence of Ku. d Cross-sectional TEM micrographs comparing interface roughness in fabricated MTJs: smooth (h = 0.1 nm) vs. orange-peel (h = 0.4 nm) interfaces. e Tunnel magnetoresistance (TMR) versus perpendicular magnetic field (H) for d = 1.2 nm at varying temperatures. f Hoffset as a dependence on the temperature with different h. Temperature-dependent offset field (Hoffset) evolution with different interface roughness.

This model systematically investigates the governing parameters of interfacial magnetic coupling. As demonstrated in Fig. 2b, under fixed geometric constraints (wavelength W = 9.5 nm, MgO thickness d = 1.2 nm, magnetic layer thickness b = 1 nm), the interfacial anisotropy constant Ku emerges as a critical coupling determinant. For interface roughness h = 0.4 nm, ΔE transitions from negative (FM) to positive (AFM) when Ku decreases below 6 × 105 J/m3. Conversely, reduced roughness (h = 0.1 nm) yields Ku-insensitive coupling with near-zero ΔE. These results establish Ku modulation as an effective control knob for coupling reversal in orange-peel coupled systems. Remarkably, Ku-mediated coupling control exhibits strong parametric selectivity, showing minimal dependence on saturation magnetization (Ms) and layer thickness variations (Supplementary Note S2).

To systematically investigate the temperature dependence of Ku, we conducted temperature-dependent magnetometry on a W(5 nm)/CoFeB(1 nm)/MgO(1.2 nm) heterostructure. The Ku values, extracted through Stoner–Wohlfarth formalism (Supplementary Note S3), demonstrate strong thermal sensitivity in this temperature regime. As shown in Fig. 2c, the Ku exhibits monotonic reduction from 300 K to 400 K. This temperature-induced Ku modulation directly enables reversible thermal control of interfacial exchange coupling, possibly providing a physical pathway for thermally activated AFM/FM transition in orange-peel coupled systems.

We first engineered MTJ films with tailored interfacial roughness using a W(5 nm)/CoFeB(1 nm)/MgO(d = 1.0, 1.2 nm)/CoFeB(1.2 nm)/W/SAF stack. Cross-sectional TEM analysis (Fig. 2d) confirmed the designed interface morphologies: the d = 1.2 nm MgO layer exhibited periodic orange-peel coupling (T = 9.5 nm, h = 0.4 nm) matching the modeling parameters in Fig. 2b, while the d = 1.0 nm interface showed atomically smooth profiles (h < 0.1 nm). The means to realize specific interface roughness has been reported in our previous work30.

To probe the coupling field of different interfaces, we performed temperature-dependent hysteresis loop measurements (h = 0.4) in Fig. 2e (hysteresis loops for device of h = 0.1 is in Supplementary Note S4). For a rough interface (h = 0.4 nm), the Hoffset first increases with temperature before undergoing a reversal (Fig. 2f), mirroring the predicted FM/AFM transition in Fig. 2b. This non-monotonic evolution arises from competing thermal effects: initial enhancement of FM coupling through activated FL and RL magnetic alignment at moderate temperatures (300–350 K), followed by decreased Ku driven AFM dominance and the reduced stray field from the SAF as heating progresses (350–400 K, Fig. 2c). In contrast, smooth interfaces (d = 1.0 nm) exhibited Hoffset reduction (Fig. 2f), consistent with Ms thermal decay from the stray field of SAF without orange-peel coupling contributions31.

To evaluate the temperature fluctuation in MTJ during writing operation, we did simulations of Joule heating in Supplementary Note S5. The write pulses elevate MTJ temperatures from 300 to 400 K, fully spanning the Ku threshold required for a coupling reversal. Using write pulse to achieve unipolar MTJ by controlling the interlayer coupling is highly possible. Other parameters of influencing temperature of the device, such as the variation of R due to spacer layer thickness and interface conductance are quantified in Supplementary Note S6.

Experimental demonstration of the unipolar MTJ

The thermal controlled coupling mechanism enables unipolar MTJ switching through precise current amplitude modulation. During write operations, an auxiliary out-of-plane field compensates SAF stray fields—a temporary requirement addressable via structural optimization. As illustrated in Fig. 3a, three distinct operational regimes emerge under increasing current density. Low J (RT) maintains state stability for reliable readout. Intermediate J increases the device temperature to T1 and writes P states through lower Ku-enhanced FM coupling (Fig. 3b). Higher J heats the device to T2 and further reduces the Ku and MSAF, activating AP state. This current-dependent thermal control successfully achieves deterministic unipolar switching.

Fig. 3: Demonstration of the unipolar MTJ.
Fig. 3: Demonstration of the unipolar MTJ.
Full size image

a Current induced switching loops of the unipolar MTJ. b Switching probabilities of the MTJ of AP to P and P to AP, respectively. c Current-induced switching under magnetic field of Hx = 180 Oe. d Schematics of writing of P and AP state at different temperature. e Current induced switching curve of the 2D-1MTJ structure at pulse duration of 50 ns and 1 ns.

The switching probability characteristics of the unipolar MTJ were systematically quantified in Fig. 3b. At RT, both AP → P and P → AP transitions exhibit zero switching probability, ensuring stable readout integrity. With increasing current density, the AP → P switching probability initiates an upward trend, achieving near-deterministic conversion (>99%) at critical current density exceeding J > 3.3 × 107 A/cm2. This trend reverses when the current density surpasses 4 × 107 A/cm2, where the AP → P probability decreases concomitant with rising P → AP switching efficiency. Ultimately, the P → AP transition approaches near 100% switching rate at current densities beyond 4.5 × 107 A/cm2. These distinct probabilistic thresholds establish well-separated operational windows for reliable read and write operations under unipolar current.

The proposed unipolar MTJ retains functional compatibility with conventional SOT-MTJ devices. Bipolar current switching measurements under an in-plane magnetic field (Hₓ = 180 Oe) reveal characteristic square hysteresis loops with counterclockwise chirality in Fig. 3c, confirming SOT-driven magnetization dynamics. The SOT significantly reduces both the effective switching energy barrier and lowers the coercivity of the MTJ. For P state writing, Joule heating induced temperature T1 shifts the hysteresis loop to the right in Fig. 3d (FM coupling), wherein the combined effects of SOT-mediated coercivity reduction drive the MTJ into the P state. Upon current pulse termination, the absence of SOT assistance results in coercivity recovery, stabilizing the P state during cooling. This mechanism operates reciprocally for AP state writing at temperature T2. One difference is that the cooling trajectory from T2 to RT traverses T1. At T1 in this situation, the P state cannot be formed due to the elevated coercivity in SOT-free conditions. As a result, the AP state can be maintained during the cooling process from T1 to RT. Thus, the SOT plays an essential role in the writing of P or AP configurations by controlling the coercivity of the device.

This SOT-enabled switching also permits ultrafast operation of MTJs across nanosecond timescales, as evidenced by successful magnetization reversal at both 50 ns and 1 ns pulse widths (Fig. 3e). Detailed analysis of the switching dynamics under 1-ns current pulses is provided in Supplementary Note S7. Simulations determining threshold pulse widths are presented in Supplementary Note S8.

Capitalizing on the unipolar current characteristic, we implemented a selection scheme through diode-MTJ integration (inset of Fig. 3e). The diodes’ unidirectional conduction property inherently restricts effective switching to negative current polarities, creating a self-selecting means that eliminates conventional transistor-based selection requirements.

Parallel writing of the unipolar MTJs

To enable parallel writing, we implemented a shared-electrode architecture where four MTJs share a common SOT channel while maintaining individual addressability. This architecture leverages VCMA effects by applying VG across the MgO barrier to locally modulate switching thresholds. As detailed in Supplementary Note S9, VG tuning enables precise control of both coercive field (Hc) and anisotropy field (HK), allowing dynamic adjustment of critical switching currents of an individual cell without crosstalk issue.

The voltage-controlled modulation of switching characteristics was investigated through VG application across the MgO barrier. Resistance-current (R-J) measurements in Fig. 4a demonstrate reversible threshold adjustments, where negative VG decreases the critical switching current while positive VG increases it. This voltage control is further quantified in the Psw measurements for P → AP transitions (Fig. 4b), revealing systematic curve shifts under VG variation. The shift of Psw created by VG enable dual-state writing at identical current densities—negative biases facilitate AP state, whereas positive biases favor P state.

Fig. 4: Parallel writing of the 5D-4MTJ structure.
Fig. 4: Parallel writing of the 5D-4MTJ structure.
Full size image

a RMTJ-J loops of the MTJ by different VG. b Psw of P to AP switching with different VG. c The schematics and PCB of 5 diodes and 4 MTJs structures (5D-4MTJ). d Experimental demonstration of parallel writing of the 5D-4MTJ structure.

Table 1 Comparation of different bit cells
Full size table

The multi-MTJ integration scheme employs a shared bottom electrode (BE) architecture, as illustrated in Fig. 4c, where four MTJ devices are interconnected via a printed circuit board (PCB). This configuration implements a 5D-4MTJ (five diodes, four MTJs) selector network that decouples read/write paths: A shared diode controls write operations while four dedicated diodes govern individual read channels. As experimentally demonstrated in Fig. 4d, localized gate voltage (VG) adjustments enable parallel writing of individual MTJs with the existence of J.

Discussions

Conventional STT/SOT-MRAM designs can write state bits (all ‘1’s or all ‘0’s) within a single clock cycle, requiring two cycles for arbitrary data patterns. Nevertheless, our approach delivers substantially greater speed advantages due to fundamental differences in circuit architecture, channel utilization, and scalability. The achievable parallelism in conventional designs is constrained by the requirement for independent write circuits - including complex row/column decoders and drivers capable of selecting individual bits.

When comparing systems with equivalent circuit complexity to write n channels, our parallel write architecture achieves significant performance gains. Conventional STT(SOT)-MRAM requires ~2t time (t = write cycle duration) to complete writing operations through n bits (n channels), yielding average speed ~2t/n. Our shared-channel 5D-4MTJ structure writes 4n bits for n channels. This results in substantially faster average operational speed (t/4n), representing at least an eightfold improvement over conventional design.

This acceleration stems from our architecture’s inherent trait: shared channels simultaneously write all MTJs for the active state, eliminating per-cell selection during write operations. This intrinsic feature removes significant write-path latency and circuit complexity inherent in conventional approaches. The demonstrated t/4n average time per bit (8× speedup) constitutes a conservative baseline from our 4-MTJ-per-channel implementation. Crucially, this parallelism scales with process technology advancement. As feature dimensions decrease, integrating additional MTJs per shared write channel becomes feasible, enabling proportionally greater speed enhancements without corresponding increases in write circuit complexity.

The concurrency-driven approach of the 5D-4MTJ architecture enables parallelized array operations, which dramatically enhances effective write speed—a four-device array written in 1 ns translates to an equivalent 250 ps per-bit write latency. Table 1 presents the simulation results for 2T-1MTJ, 2D-1MTJ and 5D-4MTJ MRAM cells (the layouts were designed in Supplementary Note S10). In comparison to the 2T-1MTJ SOT-MRAM [13], the average speed can increase by 4 times due to the parallel writing. Compared to conventional 2T-1MTJ cell, the 5D-4MTJ configuration reduces bit-cell area by 80% (from 0.0162 μm² to 0.003 μm²) by replacing area-intensive transistors with vertically stackable Schottky diodes.

The enhanced parallel writing capability of the 5D-4MTJ configuration demonstrates a 75% reduction in energy consumption per bit compared to conventional 5T-4MTJ architectures, primarily attributable to its distinct data writing mechanism. This significant reduction stems from the fundamental disparity in their operational mechanisms: the 5D-4MTJ architecture achieves parallel writing of four bits per clock cycle through optimized magnetic tunnel junction synchronization, whereas the 5T-4MTJ implementation remains constrained to serialized single-bit operations. The 5D-4MTJ architecture achieves dual optimization by simultaneously boosting processing speed and reducing energy consumption. The WER comparation between the unipolar MTJ and SOT-MTJ base line device is in Supplementary Note S11.

Methods

Sample growth and device fabrication

The film stack, consisting of W (5 nm)/CoFeB (0.9 nm, FL)/MgO (1,1.2 nm)/CoFeB (1.1 nm, RL)/W (0.6 nm)/Co (0.6 nm)/Synthetic antiferromagnetic (SAF)/Hard mask (HM) (starting from the bottom), was deposited using the Ulvac SME200 magnetron sputtering system, as shown in Fig. 2d. Following deposition, the wafers were annealed in vacuum at 350 °C for 1 h to ensure the preservation of perpendicular magnetic anisotropy and interface roughness.

The devices are manufactured on a 200 mm back-end-of-line (BEOL) platform. The MTJ and the bottom electrodes of the MTJs are patterned through deep ultraviolet lithography and etched using inductively coupled plasma etching techniques. The MTJ is encapsulated with layers of SiN and SiO2. The top electrode is composed of Al/TiN.

Electrical measurement of MTJs

Electrical transport measurements are performed using a magnetic field probe station supplied by East Changing Technologies. Write pulses down to 0.8 ns is generated from the PG 1072 pulse generator. The resistance of the MTJ is measured by Keithley 2602 after the write pulses. VG applied on the junctions are also applied by Keithley 2602. It was applied before the write pulse. Attenuators were connected to the read channel of the MTJs to compensate the voltage induced by the write pulse.