Who has the greatest hope in the new-type storage field?
Storage technology is at the core of modern computing systems. From basic data storage to more complex tasks, such as "in-memory computing" for artificial intelligence (AI) and machine learning (ML) applications, all rely on their support. These technologies were initially only used for data retention, but now they are gradually evolving to adapt to new computing paradigms, such as "in-memory computing," which involves direct data processing within the storage array. This evolution has significantly improved computing efficiency as it reduces data transfer between the processor and the memory, thereby increasing speed and reducing energy consumption - which is particularly crucial for high-load tasks like AI and ML. It is these demanding performance requirements that are driving technological innovation to break through the limitations of the traditional complementary metal-oxide-semiconductor (CMOS) paradigm.
Emerging non-volatile memories (eNVMs) represent a promising class of technologies that can be used to replace or enhance traditional volatile memories (such as random access memory RAM). Unlike RAM, which loses data when the power is turned off, eNVMs can maintain data integrity even when the power is off or the system is shut down. This article reviews various emerging storage materials and device architectures, including resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FeRAM), and phase-change memory (PCM). In addition, it explores new types of eNVMs based on two-dimensional materials and organic materials, and discusses the transition from traditional digital computing to neuromorphic computing, and how this transition brings new opportunities to address the technological bottlenecks faced by artificial intelligence in accelerating scientific research discoveries. This article systematically analyzes the current technological progress, development trajectory, and the main challenges that still need to be overcome.
The Role of Non-Volatile Memories in the Post-CMOS Era
In the post-CMOS microelectronics era, a widely recognized challenge is how to break through the limitations of the von Neumann computing architecture (see Figure 1). There is an urgent need for a new type of memory that combines multiple advantages, including compatibility with existing CMOS process flows and the ability to overcome the scaling bottlenecks of static random access memory (SRAM) and flash memory. Storage technologies with these characteristics can be applied to independent storage or embedded storage in analog and digital processing. According to the 2022 report of the International Roadmap for Devices and Systems (IRDS), such technologies are expected to trigger a revolution in computing architecture.
The research on non-volatile memories can be traced back to the charge storage devices of the 1960s. The research lasted for several decades until the embedded semiconductor storage technology was scaled down to the 28-nanometer node in 2010. However, subsequent further miniaturization was hindered due to charge leakage issues. One of the key advantages of non-volatile memories is their data retention ability (degree of non-volatility), which is usually measured by the length of time the data can be retained. Currently, due to its high technological maturity, excellent optimization, and wide commercial applications, flash memory is regarded as the benchmark technology for non-volatile storage.
Figure 1 (a) Visual representation of the development history of non-volatile storage; (b) Timeline of technological development; (c) Classification of emerging memory devices by maturity; (d) Prediction of circuit architecture by 2035, characterized by the integration of multiple emerging storage technologies selected to meet the functional requirements of the chip.
Typical storage technologies have now reached commercial availability, and a complete scientific, technological, and systematic knowledge system has been established in the literature. Since it is difficult for charge-based memories to achieve nanoscale layer thicknesses, the current technological focus has shifted to the three-dimensional stacking structure of NAND flash memory and various "emerging" memories. Figure 1c shows six major categories of emerging storage technologies, ranked from high to low maturity: new magnetic storage (MRAM), ferroelectric storage (FeRAM), and oxide-based resistive storage (ReRAM). These technologies have demonstrated good characteristics and are ready to enter the commercial verification stage. The following technologies are still in the early stages of development but have great potential for technological breakthroughs: phase-change memory (PCM or PRAM), conductive bridge memory (CBRAM), two-dimensional material memory (2D RAM), and organic and molecular memory. Some molecular-level storage technologies, such as Mott memory and DNA data storage, are still in the preliminary exploration stage.
Due to their unique characteristics such as non-volatility, byte addressing, high density, high scalability, and near-zero standby power consumption, storage-based computing and processing will play an irreplaceable role in future computing systems. With the rapid development of neuromorphic memories, combining them with emerging memories in the future is expected to completely transform the computing architecture, improving system performance, energy efficiency, and processing capabilities, applicable to all levels from storage systems to edge and cloud environments, database systems, and even blockchain decentralized applications.
Diversity and Advantages of Storage Technologies
The diversity of emerging storage technologies (such as ferroelectric memory FeRAM, redox resistive memory ReRAM, magnetic memory MRAM, phase-change memory PCM, and organic and molecular memory OMRAM) provides a variety of options for specific application requirements, allowing designers to configure flexibly according to the required specifications and operating environment. Each technology has unique advantages, such as high durability, excellent energy efficiency, and the ability to adapt to specific environments or tasks.
Research on non-volatile memories for high-temperature environments addresses the need for reliable operation under extreme conditions. This type of research fills an important gap in the existing technology market and expands the potential applications of memories in harsh environments. For example, through innovation in material selection and improvement in manufacturing precision, storage devices can operate stably even under extreme conditions such as high temperature and high radiation, which is crucial for industries such as aerospace and geothermal exploration.
In the rapidly developing field of storage technology, two-dimensional (2D) materials are emerging as a promising new path due to their unique physical properties and good scalability. These materials, with their atomic-level designability and compatibility with existing technologies, are expected to revolutionize storage devices. The characteristics of 2D materials, such as atomic-level thickness and structural design flexibility, enable them to achieve faster and more energy-efficient memories and integrate seamlessly with current electronic technologies, improving the performance of the entire system. With the continuous progress of material synthesis and transfer processes, the large-scale application of 2D materials is gradually becoming a reality, indicating a new stage in the development of next-generation storage technologies that can meet the future computing and data storage needs.
Different from traditional storage technologies, ReRAM and synaptic RAM support "in-memory computing," are non-volatile, and can achieve low-latency and low-energy data processing. They can directly perform analog multiply-accumulate operations within the storage array, eliminating the energy-intensive "storage - processor" data transfer bottleneck in the traditional von Neumann architecture. This makes them very suitable for edge computing systems, especially in application scenarios with key requirements such as real-time inference, low power consumption, and compact design.
These storage technologies are particularly suitable for brain-inspired computing and adaptive systems. Synaptic RAM, inspired by biological synapses, can implement learning mechanisms such as "spike-timing-dependent plasticity (STDP)," enabling hardware-based learning and real-time response in dynamic environments. This ability is crucial for self-learning Internet of Things (IoT) devices, allowing them to sense and adapt to new environments without continuous cloud connection. In addition, the non-volatile characteristics of xRAM (such as ReRAM, FeRAM, MRAM) or PCM enable the IoT system to maintain its operating state after a power outage, enhancing the system's reliability and enabling the instant wake-up function in environments with limited or intermittent power supply.
With the increasing demand for intelligent and decentralized systems, xRAM and synaptic RAM are becoming important technological paths for realizing scalable, low-power, and highly robust computing platforms. They have advantages such as high density, support for 3D integration, and the ability to be monolithically integrated with CMOS circuits, playing a central role in the evolution of AI and IoT hardware architectures. These technologies support the vision of "distributed intelligence," enabling computing systems to operate seamlessly, autonomously, and with context awareness in a wide range of scenarios, including intelligent sensors, edge analytics, and more.
Figure 2. Schematic diagram of the roadmap for three-dimensional integrated brain-inspired hardware using two-dimensional materials. Reprinted from Kim, S.J., et al., "Three-dimensional integrated brain-inspired hardware based on two-dimensional materials," NPJ 2D Materials and Applications, Volume 8, Page 70 (2024), with permission from Nature.
Non-Volatile Memories on Flexible Substrates: A Review of Cutting-Edge Technologies
The integration of non-volatile memory (NVM) technology onto flexible substrates has received extensive attention in recent years, mainly driven by emerging applications such as wearable electronic devices, soft robots, and distributed Internet of Things (IoT) systems. These systems not only require the memory to retain data when the power is off but also to withstand mechanical deformations such as bending, stretching, and twisting. Among various NVM technologies, ReRAM and FeRAM are the most advanced on flexible platforms.
Due to its simple metal-insulator-metal structure and good compatibility with low-temperature processes, ReRAM has demonstrated excellent mechanical durability and data retention ability on polymer substrates (such as PET and polyimide). FeRAM (especially devices based on P(VDF-TrFE) or doped HfO₂) has low-voltage switching and stable polarization states, maintaining reliability after thousands of mechanical cycles. Although more challenging, flexible PCM and MRAM based on organic materials or two-dimensional magnetic materials have also entered the preliminary research and demonstration stage.
Organic and molecular storage technologies are also developing rapidly, showing great potential in edge AI computing and bio-inspired devices. Organic materials have tunable molecular structures and corresponding electrical, optical, thermal, and chemical properties, which can replace traditional memristors in some neuromorphic computing algorithms. The volatility and dynamic electrical characteristics of these materials and devices enable them to simulate the response functions of biological neurons and synapses, including spike-timing-dependent plasticity (STDP), spike-frequency-dependent plasticity, and short-term and long-term plasticity.
In recent years, with the progress of materials and manufacturing technologies (such as inkjet printing, transfer technology, and room-temperature deposition processes), it has become possible to directly fabricate non-volatile memories on plastic substrates without damaging their mechanical integrity. Hybrid material systems based on two-dimensional materials and nanostructured dielectrics have further improved the device performance under stress conditions. Despite these achievements, challenges still remain, such as how to achieve long-term mechanical reliability, maintain data under bending conditions, and how to integrate flexible NVM with logic and sensing elements to build a complete flexible system. Future research will focus on monolithic integration, additive manufacturing of the entire storage array, and the development of robust packaging methods to achieve stable operation in dynamic environments.
As these obstacles are gradually overcome, flexible non-volatile memories (NVM) will become the core of adaptable and conformable electronic devices. These flexible neuromorphic integrated circuits can be easily combined with organic and molecular storage technologies. In such applications, neural network models can be trained on a central high-performance AI chip, and then the models can be deployed on flexible neuromorphic integrated circuits close to biochemical sensors for local processing. This model is particularly suitable for wearable or implantable biomedical devices, extending AI computing capabilities to distributed computing and sensing systems.
The key knowledge gaps restricting the application of emerging storage technologies include: (a) eliminating material impurities and inhomogeneities, which can lead to poor write durability and short lifespan; (b) understanding the impact of nonlinear dynamic behavior and analog noise on memory in computing and processing; (c) shortening the time required for memory reprogramming and reducing the energy consumption of the write process.
Manufacturing and Use of Emerging Memories
The manufacturing of emerging storage materials often requires deposition processes in an ultra-high vacuum environment. Precision manufacturing facilities can ensure high-precision fabrication of storage devices and avoid the introduction of contaminants. The use of integrated deposition and material characterization tools can achieve seamless connection at different manufacturing stages, ensuring the purity and functional integrity of the materials throughout the process. The use of these high-end tools not only improves the performance of the memory but also significantly extends the working life and operational reliability of the device.
Understanding the basic properties of materials used in storage devices is crucial. Mistakes made in the early development stage may cause serious performance problems in subsequent stages. When using emerging storage technologies for storage applications, fast and low-energy switching needs to be achieved to efficiently write and read data while consuming minimal energy. In contrast, the storage device must remain stable after programming to ensure reliable in-memory computing. Once programmed, the device will be read multiple times during the computing process. Therefore, demonstrating good repeatability and durability during these read processes is essential for accurate processing. Improvements in characterization techniques, such as in-situ measurement and detailed statistical analysis across device regions, help to better understand the behavior and performance of materials. This will result in more consistent and reliable storage devices, which is crucial for their application in high-risk industries.
The development of multi-channel test systems for emerging storage devices has promoted a more efficient and accurate testing process. These test systems are essential for designing and developing energy-efficient hardware that can meet the computing requirements of AI models, significantly reducing the energy consumption of these technologies. Each advancement addresses specific challenges in the field of storage technology, laying the foundation for future innovations that may redefine the way storage devices are integrated and used in various technological platforms. As these technologies evolve, they continuously push the boundaries of computing power, heralding a new era of higher speed, greater efficiency, and stronger reliability. The development of storage technology not only reveals potential innovations and technological breakthroughs but also highlights the challenges that need to be addressed urgently. These challenges can be grouped into five categories: (a) material synthesis, manufacturing precision, and characterization; (b) device scalability, lifespan, and repeatability; (c) multi-modal characterization of materials and devices; (d) device interconnectivity and compatibility with existing and new CMOS technologies; and (e) packaging and heterogeneous integration.
In terms of material synthesis, it is challenging to select materials that can withstand extreme conditions (such as high temperature and high density) and meet the computing requirements of AI and ML applications. In addition, the composition of the materials must be precisely defined to ensure their stability and functionality. Developing AI/ML methods to select new composite materials can accelerate the synthesis process. Achieving high-quality materials requires precision control in the manufacturing process, which is crucial to avoid defects that may reduce storage performance. This includes maintaining an ultra-high vacuum state during thin-film deposition to prevent contamination. High-quality materials can provide stability, repeatability, and scalability to the device.
Due to their unique electrical, mechanical, and thermal properties at the atomic scale, two-dimensional (2D) materials are crucial for the development of emerging non-volatile memory (eNVM) technologies. Their atomic-level thinness allows them to break through the scaling limits of traditional semiconductors, enabling high-density and low-power storage integration. Materials such as graphene, transition metal dichalcogenides (such as MoS₂, WS₂), and hexagonal boron nitride offer high carrier mobility, tunable bandgaps, and switching characteristics achievable through defect engineering, making them very suitable for resistive switching, ferroelectric behavior, and charge trapping mechanisms in emerging non-volatile memory (eNVM) devices. In addition, their mechanical flexibility and chemical stability make them suitable for flexible and wearable electronics, where traditional materials perform poorly. These characteristics make it possible to realize next-generation storage devices with fast switching speed, excellent durability, and retention ability, which are essential for in-memory computing and AI applications. Although this technology has great potential in storage devices, 2D materials also bring a series of unique challenges. Stable production of large-area, high-quality single-crystal 2D materials is an important technological obstacle that must be overcome for their widespread application. These materials are susceptible to environmental factors including oxygen and moisture, which may damage their performance. Developing effective packaging and protection strategies is crucial for the practical use of these materials. Integrating 2D materials with existing manufacturing processes (especially CMOS technology) also requires innovation in low-temperature growth technology and non-destructive transfer methods.
To meet the high standards required for commercial viability and manufacturing precision, advanced characterization techniques must be adopted. Achieving repeatability of characterization results across different device regions is crucial but extremely challenging due to the variability of material behavior and defects. Real-time in-situ measurement is essential for understanding the internal dynamic interactions during device operation, but it also poses significant technical challenges, especially in modifying tools such as transmission electron microscopy (TEM) for real-time operation with a charged bias. Developing storage technologies that can operate at extreme temperatures also faces important challenges: