STT-MRAM (Spin Transfer Torque MRAM) is distinguished emerging memory that combines high speed and low power operation. Originally developed for high-capacity memory applications, it has recently been mass-produced as embedded-MRAM with high reliability, surpassing the limitations of embedded-flash memory by various foundries. Although it has been primarily used in memory applications so far, recent advancements in new operating mechanisms and diverse applications have generated great expectations for the future. The foundation of these new technologies and applications lies in the Magnetic Tunnel Junction (MTJ) technology, which converts its magnetic characteristics into electrical signals, specifically data “0” and “1”. Therefore, this article introduces and explores various memory and computing technologies that utilize MTJ, highlighting their prospects and future developments.

The first is the traditional non-volatile memory mentioned in the introduction, namely STT-MRAM technology. MTJ is a device with a degree of freedom in which the switching conditions, switching speed, and reliability are controlled by the magnetic/electrical characteristics and physical size (=CD) of pinned, free, and barrier layers. As a result, STT-MRAM has been developed and mass-produced in various fields such as high-capacity memory, high-speed cache memory, and highly reliable automotive memory through MTJ stack design. In the future the need for high-speed, low-power, and high-reliability STT-MRAM is expected to continue, leading to requirement of complementary and integrated technologies in process, device, and circuit design.

Secondly, SOT-MRAM (Spin Orbit Torque MRAM) is receiving significant attention from academia and industry as a potential solution to overcome the speed and reliability limitations of STT-MRAM. SOT-MRAM has the advantage of separating the read and write paths and using spin currents generated in the channel to rapidly switch the magnetic layer with sub-nanosecond speed. Recently, there has been active research conducted worldwide, to the extent that SOT-MRAM accounts for half or two thirds of MRAM papers presented at major semiconductor conferences such as IEDM, VLSI, and IMW. Although there are still challenges to be addressed, such as high write currents and external magnetic fields required, researchers around the world are actively pursuing various approaches, including material and structural methods, to overcome these challenges.

Third, the recent focus of researchers in the field of emerging memory is on next-generation computing research known as In-memory computing (IMC) or Computing In Memory (CiM), utilizing various non-volatile memory (NVM) devices such as resistive switching material, phase change material, and ferroelectric. These exhibit analog switching behavior, making them suitable for Analog-CiM (A-CiM) applications. In contrast to these, STT-MRAM and SOT-MRAM has basically only binary states of P and AP, making the formation of multi-level states very difficult and limiting its performance and application areas in CiM. However, many papers have been published on the use of STT-MRAM or SOT-MRAM-based accelerators and neural networks in fields that require high switching accuracy, low write energy, and high reliability. Furthermore, research on the application of CiM-based AI accelerators to neuromorphic computing neurons or synapse components is actively underway. 

Finally, in addition to neuromorphic computing, probabilistic computing (p-bit computing) can be discussed as a next-generation computing technology that overcomes the limitations of conventional von Neumann computers. P-bit computing utilizes the probabilistic switching of MTJ, where the operating conditions (voltage, current, switching time, etc.) determine the switching probability. This is in contrast to deterministic switching used in conventional memory devices such as DRAM or NAND. P-bit computing intentionally creates imperfect switching by adjusting the magnitude of thermal stability (Δ), unlike memory components that require perfect switching. The von Neumann computer is highly efficient in operations such as arithmetic and bitwise operations, but it is known to be inefficient in areas where precise logic or rules are difficult to define or computationally challenging problems. On the other hand, p-bit computing can efficiently solve problems that involve finding optimal solutions among numerous possibilities. 

Starting from conventional MRAM and extending to STT-MRAM, SOT-MRAM, and next-generation computing technologies, MTJ has created various opportunities and will continue to do so in the future. These technologies do not demand a single characteristic but rather require different MTJ properties depending on the application. Those who possess powerful MTJ technology (process, device, circuit design, etc.) will be able to seize great opportunities in the era of next-generation memory and computing. Currently, the focus of semiconductor devices utilizing MTJ is still memory, and the key technologies for high-capacity memory will provide additional opportunities for exploring new application targets. Based on the diverse world-class high-capacity memory technologies accumulated through the development of 4G STT-MRAM for DRAM and the 1S1M cross-point for SCM, SK Hynix inc. is preparing for the upcoming future.