In recent years, non-volatile memory (NVM) technology has been rapidly advancing. With the rise of emerging applications such as artificial intelligence, autonomous driving, and the Internet of Things, traditional storage systems are facing challenges in speed, energy consumption, and stability. In response to the demands for speed, efficiency, and reliability, various new memory technologies including ReRAM, PCM, FeRAM, and MRAM have entered research and validation stages, aiming to emerge in the "post-DRAM era." In this context, magnetic random-access memory (MRAM), which offers high speed, low power consumption, and non-volatility, is considered one of the most promising general-purpose storage solutions. Reports indicate that a multinational research team from National Yang Ming Chiao Tung University, Taiwan Semiconductor Manufacturing (TSM), and the Industrial Technology Research Institute has made a significant breakthrough in MRAM technology. They successfully developed a spin-orbit torque magnetic random-access memory (SOT-MRAM) based on β-phase tungsten material, achieving impressive performance metrics: data switching in just 1 nanosecond, data retention for over 10 years, and a tunneling magnetoresistance (TMR) ratio of up to 146%. This achievement, published in the journal Nature Electronics, paves the way for the industrial application of the next generation of high-speed, low-power storage technology.
The transformation in storage technology is driven by the current reliance of computing systems on a hierarchy of storage comprised of SRAM, DRAM, and flash memory. However, with technology nodes breaking below 10 nanometers, these traditional charge-storage-based methods face severe challenges including scalability limits, difficulties in performance enhancement, increased read/write interference issues, and decreasing reliability. As artificial intelligence and edge computing rapidly evolve, there is a rising demand for memory that not only possesses the high-speed response of DRAM but also the non-volatile features of flash memory, all while significantly reducing power consumption. In this context, emerging non-volatile storage technologies have emerged. Besides SOT-MRAM, this includes spin-transfer torque magnetic random-access memory (STT-MRAM), phase-change memory (PCM), resistive random-access memory (RRAM), and ferroelectric random-access memory (FeRAM). These technologies are characterized by non-volatility, low latency, and low power consumption, and can be integrated with existing CMOS semiconductor processes, enabling the development of new computing architectures.
In comparison, DRAM has a latency of about 14 milliseconds, while 3D TLC NAND has read latency ranging from 50 to 100 microseconds. The new SOT-MRAM achieves switching speeds at the nanosecond level, nearly rivaling SRAM, while also retaining non-volatility—meaning data is not lost even when power is cut off.
The unique advantages of SOT-MRAM are attributed to its distinct operating principle and technical benefits. It utilizes materials with strong spin-orbit coupling to generate spin-orbit torques (SOTs), facilitating the magnetization reversal of magnetic tunnel junctions' nano-magnets, thereby accomplishing data writing and erasing. Compared to other storage technologies, SOT-MRAM holds three core advantages:
1. High-speed writing: The spin-orbit torque effect enables magnetization reversal within nanoseconds, significantly faster than traditional magnetic field-driving methods. 2. High energy efficiency: Its three-terminal architecture completely separates read and write current paths, effectively addressing STT-MRAM durability issues and the resistance limitations of magnetic tunnel junctions, while substantially reducing power consumption. 3. High reliability: Independent read and write operations substantially improve the device's durability, allowing for more read/write cycles while maintaining excellent long-term data retention.
These advantages position SOT-MRAM as a potential replacement for high-speed SRAM, aiming to become a core storage component of next-generation computing systems.
Despite the clear theoretical advantages of SOT-MRAM, a critical technical bottleneck must be addressed for its industrial application: the thermal stability of spin-orbit coupling materials. Tungsten, due to its strong spin-orbit coupling characteristics, is an ideal candidate material for SOT-MRAM. The transition of tungsten to thermodynamically stable α-phase tungsten during semiconductor manufacturing processes, typically requiring thermal treatments at 400°C for several hours, presents a significant challenge. This phase change is detrimental—α-phase tungsten has a spin Hall angle of only about -0.01, drastically reducing spin-orbit torque flipping efficiency and severely degrading device performance.
The research team's breakthrough solution involves inserting ultra-thin cobalt layers within the tungsten layer to form a composite structure. Specifically, they divided a 6.6 nanometer thick tungsten layer into four segments, inserting only 0.14 nanometer thick cobalt layers between them—this thickness is smaller than a single atom layer of cobalt, resulting in cobalt presenting a discontinuous distribution. This clever design serves a dual purpose: the cobalt layers act as diffusion barrier layers, suppressing atomic diffusion within the tungsten; while the mixing effect between cobalt and tungsten consumes thermal budget, thereby delaying the occurrence of phase transitions.
Exciting experimental validations were achieved: this composite tungsten structure can maintain phase stability at 400°C for up to 10 hours and withstand temperatures of 700°C for 30 minutes, while traditional single-layer tungsten undergoes a phase change after only 10 minutes of annealing at 400°C. Through transmission electron microscopy, X-ray diffraction, and nano-diffraction testing at the Taiwan Photon Source, researchers confirmed the stability of β-phase tungsten. More importantly, this composite structure resolves thermal stability issues while maintaining excellent spin conversion efficiency. Using spin torque ferromagnetic resonance and harmonic Hall resistance measurements, the team found that the spin Hall conductivity of the composite tungsten film is approximately 4500 Ω⁻¹·cm⁻¹, with a damping-type torque efficiency of about 0.61, parameters ensuring high-performance magnetization flipping.
Comprehensive validation of performance shows that theoretical breakthroughs must be substantiated through device verification. The research team successfully fabricated a 64Kb SOT-MRAM prototype array based on the composite tungsten film scheme and completed comprehensive performance testing and validation under near-real application conditions. In terms of switching speed, the device achieved spin-orbit torque flipping velocities in the nanosecond range, with performance nearly rivaling SRAM and far exceeding that of DRAM and flash memory. Statistical tests of 8000 devices showed highly uniform flipping behaviors, with an intrinsic flipping current density of only 34.1 mA/cm² under long pulse (10 nanoseconds) conditions, showcasing outstanding stability and repeatability.
Data retention capabilities were also impressive. Based on cumulative distribution function (CDF) estimates, the thermal stability parameter (Δ) of the devices is approximately 116, indicating a data retention time that exceeds 10 years, fully meeting the strict requirements for non-volatile storage. In TMR testing, the devices achieved TMR values as high as 146%, indicating a high-quality interface between MgO and Co₄₀Fe₄₀B₂₀, providing robust assurance for stable read margin and reliable process windows.
In terms of energy consumption, the three-terminal structure enables complete independence of read and write operations, fundamentally reducing energy usage, making it particularly suitable for power-sensitive applications such as edge computing and mobile terminals. Furthermore, thanks to the involvement of the TSM research team, the entire design has been optimized for existing semiconductor backend processes from the very beginning, ensuring outstanding process compatibility and paving the way for future mass production.
Notably, the research team also achieved X-type flipping without an external magnetic field. This achievement is attributed to the symmetry-breaking effect within the composite tungsten material, further simplifying device structure while enhancing integration and design flexibility, opening new avenues for the engineering applications of SOT-MRAM.
The implications of this research go beyond mere technological breakthroughs in the laboratory; it points towards new directions for the development of the entire storage industry. Unlike many new storage technologies that remain at the concept verification stage, the SOT-MRAM based on composite tungsten has been designed from the outset with process compatibility and manufacturability in mind. The research team has successfully fabricated a 64Kb array and plans to further expand to megabit (Mb) level integration while reducing write energy consumption to sub-picojoule levels.
In applications involving AI and edge computing, SOT-MRAM presents distinct advantages. High-frequency data access during AI training and inference processes is a major source of energy consumption, and SOT-MRAM, with its features of high speed, non-volatility, and low power consumption, can serve as cache for AI accelerators, significantly lowering system energy demand. In edge devices, its non-volatility allows for quick power cycling with no loss of data, which is particularly beneficial for battery-operated IoT devices.
Moreover, the advent of SOT-MRAM could drive a restructuring of the storage hierarchy system. The traditional "SRAM cache - DRAM main storage - flash external storage" three-tier architecture may face transformation, as SOT-MRAM is expected to fill the performance gap between SRAM and DRAM, or even replace one of them in certain applications, thereby simplifying the architecture and enhancing system efficiency.
At the materials science level, the proposed "composite layer stabilizing metastable phases" strategy not only applies to tungsten but also provides new insights for the phase stability studies of other functional materials. The team plans to explore novel oxides and two-dimensional interfacial materials to enhance overall performance and reliability further.
More broadly, this breakthrough could spur innovations in computing architectures. The high-speed and low-power features of SOT-MRAM make new architectures such as in-memory computing more feasible, providing new pathways to overcoming the "memory wall" bottleneck tied to traditional von Neumann structures.
In conclusion, the composite tungsten-based SOT-MRAM cleverly addresses the thermal stability challenges of β-phase tungsten, achieving a perfect blend of nanosecond switching and long data retention. This is not merely an academic achievement but a core technological reserve prepared for the next generation of computing systems. For the research team, their goal extends beyond demonstrating superb laboratory performance—they aim to showcase, through system-level validation, how MRAM can significantly reduce overall power consumption in practical applications, driving technological innovations in AI, edge computing, and mobile devices. As integration progresses from kilobits to megabits, we have reasonable grounds to anticipate that this new type of memory will soon be integrated into our smart devices, heralding a new era of storage technology.