Memory is one of the most critical components of modern information systems and has formed an over $100 billion market consisting mainly of DRAM and NAND Flash. With the advent of the era of smart connectivity for everything, emerging application scenarios such as artificial intelligence and smart cars are putting forward higher performance requirements for storage, prompting the rapid development of new memories and influencing the future memory market landscape.
What exactly does new memory mean, what are the technical principles, what is the competitive landscape, and what are the future prospects for development?
We focus on ReRAM (Resistive RAM) among the new memories and analyze its technology, application scenarios, and modes to make the following judgments.
1. the era of everything smart needs new memory with stronger performance in terms of speed, power consumption, and capacity.
2. comparing the four new types of memory, ReRAM has obvious advantages in terms of density, process, cost, and yield.
3. AIoT, smart cars, data centers, and AI computing (storage and computing in one) will be important development opportunities for ReRAM.
4. the IDM model is the best choice for ReRAM vendors.
The new memory is the future choice
Memory is the vane and largest market segment of the semiconductor industry, accounting for about one-third of the semiconductor industry. The arrival of the smart era will cause a new round of explosions in the storage industry.
According to YOLE, memory has become the fastest-growing segment of the semiconductor industry since 2019, and the overall market space will grow from $111 billion in 2019 to $185 billion in 2025, with a compound annual growth rate of 9%. Among the sub-segments, the new memory market will grow the fastest, from $500 million to $4 billion, with a CAGR of 42% and huge growth potential.
Global Memory Market Size and Growth Rate (Source: YOLE)
Memory can be divided into two categories according to whether power failure can save data.
The first type of volatile memory is represented by dynamic random access memory (DRAM) and static random access memory (SRAM) volatile memory, both of which have high read and write speeds. Among them, SRAM is faster than DRAM but less dense than DRAM because a DRAM memory cell requires only one transistor and a small capacitor, while each SRAM cell requires four to six transistors. Their common disadvantage is their low capacity and high cost, and they are generally used as main memory and cache, respectively.
The second type of non-volatile memory includes the traditional memory represented by NOR FLASH and NAND FLASH and four new types of memory.
In addition to FRAM, the new memories are all “0” and “1” data storage by changing the resistance value, and all four new memories are non-volatile and can save data after power failure. The four new memories are non-volatile and can save data after power failure.
Advantages of ReRAM
Currently, there are four main types of new memories.
Phase change memory (PCM), represented by the 3D Xpoint jointly developed by Intel and Micron.
Ferroelectric memory (FeRAM), represented by Ramtron and Symetrix.
Magnetic memory (MRAM), represented by Everspin (USA).
Resistive memory (ReRAM), represented by Panasonic, Crossbar, and Xindan Semiconductor.
Comparison of 4 new memory parameters (Source: Objective Analysis)
1. Phase-change memory (PCM or PCRAM)
The principle of PCM is to change the phase change material between a low resistance crystalline (conductive) state and high resistance non-crystalline (non-conductive) state by changing the temperature.
Although PCM has improved read/write speed compared to NAND Flash, its cooling process after RESET requires high thermal conductivity, which will bring higher power consumption, and because its storage principle uses temperature to realize the resistance change of phase change material, it is very sensitive to temperature and cannot be used in wide temperature scenarios. Second, in order to make the phase change material compatible with the CMOS process, PCM must adopt a multi-layer structure, so the storage density is too low to replace NAND Flash in terms of capacity.
Intel and Samsung produced the first commercial PCM chip in 2006. in 2015, Intel and Micron collaborated to develop a storage technology called 3D XPoint, which is also a type of PCM. in 2018, the two sides ended their joint development work, and in March 2021, Micron announced that it had stopped all 3D XPoint technology-based In March 2021, Micron announced that it would cease all further development of products based on 3D XPoint technology.
2. Ferroelectric Memory (FRAM or FeRAM)
FRAM does not use ferroelectric materials but is only named because the storage mechanism resembles the hysteresis behavior of ferromagnetic storage. the voltage-current relationship of FRAM crystal materials has characteristic hysteresis loops that can be used for storage.
FRAM has the advantage of fast read/write speed and good lifetime, but its memory cell is based on dual transistors, dual resistor cells, cell size is at least twice as large as DRAM, storage density is limited, and the cost is high. And its reads are destructive and must be offset by subsequent writes after each read to restore the content of that bit to its original state.
In terms of materials, the current ferroelectric crystal materials PZT (lead zirconate titanate) and SBT (strontium bismuth tantalate) suffer from fatigue degradation and environmental pollution, and perfect commercialization has yet to be found.
Currently, Ramtron (owned by Cypress) and Symetrix are leading the development of FRAM.
3. Magnetic Memory (MRAM)
The current mainstream MRAM technology is STT MRAM, which uses the “giant magnetoresistance effect” of the tunneling layer to read the bit cell.
Although STT MRAM performance is better, the critical current density and power consumption still need to be further reduced, the current MRAM storage cell size is still large and does not support stacking, the process is more complex, large-scale manufacturing is difficult to ensure uniformity, storage capacity and yield climb slowly. Until further breakthroughs are made in the process, MRAM products are mainly suitable for special applications with low capacity requirements, as well as the emerging IoT embedded storage field.
Commercially, Everspin has partnered with Global Foundries and UMC has partnered with Avalanche Technology to promote STT-MRAM.
4. Resistive Memory (ReRAM or RRAM)
Resistive RAM is a non-volatile memory based on the reversible conversion between high and low resistance states by the resistance of a non-conductive material under the action of an applied electric field. ReRAM includes many different technology classes such as OxRAM (Oxygen Vacancy Memories), CBRAM (Conductive Bridge Memories), Metal Ion Memories, and Carbon Nano-tubes.)
The cell area of ReRAM is extremely small, up to 4F², and the read/write speed is 1000 times faster than NAND FLASH, while the power consumption is reduced by 15 times.
The ReRAM process is also much simpler. Crossbar and Xinyuan Semiconductor, for example, use CMOS-friendly materials, are able to use standard CMOS processes and equipment, are non-polluting to the production line, and have low overall manufacturing costs, making it easy for semiconductor foundries to have ReRAM production and manufacturing capabilities, which is a great advantage for mass production and commercialization promotion.
ReRAM memory has significant advantages over existing new memories in terms of density, energy efficiency ratio, cost, process, and yield.
ReRAM has four major development opportunities: AIoT, smart cars, data centers, and AI
1.AIoT
AIoT refers to the integration of artificial intelligence technology and the Internet of Things in practical applications on the ground. According to Ariadne Consulting data, the total output value of China’s AIoT industry is 380.8 billion yuan in 2019 and is expected to reach 750.9 billion yuan in 2022, with a compound annual growth rate of 25.4%.
AIoT requires real-time interaction of data, so it requires not only low power consumption of memory but also high read/write and low latency. The current NOR Flash storage has low density, small capacity, and high power consumption, which cannot achieve high write speed. ReRAM, on the other hand, can increase the write speed by 1000 times with guaranteed read performance, while achieving higher storage density and very low power consumption, and will be the best choice to replace NOR Flash as the memory in the era of smart connectivity for everything.
As people and things interact with more and more information, a lot of private information will be stored and recorded, IoT brings convenience in life, but also brings potential data security risks, attacks against IoT can even be passed through the device to bring unimaginable damage in real life. AIoT applications increasingly require security attributes. PUF (Physically Unclonable Function) + new memory chip is expected to become the mainstream solution to the storage and security problems of smart devices.
PUF is a kind of chip in the semiconductor production process of process volatility to generate chip unique function, can do a core of a secret, can be called “chip fingerprint”. Currently, Xin Yuan Semiconductor has designed a PUF chip based on ReRAM memory, which can have both storage and security functions.
2. Smart car
Automotive electronics can be divided into body control systems (ECU), security systems, entertainment devices, chassis control, advanced driver assistance systems (ADAS), etc., which require semiconductor devices to achieve the relevant functions, including memory, sensors, optoelectronic devices, RF devices, power devices, etc.
According to Counterpoint Research, the future storage capacity of a single-vehicle will reach 2TB-11TB, and a L4/L5 level self-driving car needs at least 74GB DRAM and 1TB NAND. according to the IHS forecast, the global automotive storage IC market size is about $8.3 billion in 2025.
Smart cars require more than just temperature and reliability for memory. The control system needs intelligent real-time decision making; ADAS system generates a lot of image data all the time; the entertainment system needs to be more intelligent to enhance user experience; energy consumption is also a key factor for smart cars …… These all require memory with a large amount of data real-time throughput capability to ensure storage stability and high energy efficiency ratio.
Traditional NOR Flash cannot meet the requirements of future smart cars for read/write speed (especially XIP program execution efficiency); NAND Flash is difficult to achieve XIP on-chip program execution and extremely slow; DRAM and SRAM have limited capacity and data will be lost in a power failure. Among the new memories, ReRAM not only meets the requirements of high read/write speed and storage density but also reduces latency by 1000 times, which can meet the high real-time data throughput of future smart driving.
In terms of security, ReRAM has wide and mild reliability. The future is expected to appear high-performance, high integration, high stability, and low power consumption of the vehicle specification ReRAM memory.
3. Data Center
In the AI era, data is exploding and more and more data will be processed in the cloud. According to Cisco forecast, there will be 1,327EB data stored in data centers worldwide in 2021, with a 6-year CAGR rate of 41%.
The explosion in data volume has spawned new incremental markets and performance requirements for memory, and according to SUMCO forecasts, data center demand for SSD storage will grow at a compound rate of 46% between 2019 and 2023. However, data center memory performance is currently not evolving at a rate that can keep up with computing demand, and power consumption remains one of the highest cost factors in data centers.
Traditional mechanical hard drives have a long life and low cost, but the read/write speed is extremely low and the heat and noise are obvious. DRAM is faster but volatile memory, which cannot save data in a power outage and is extremely costly, and cannot be used as mass storage data. NAND read and write speed is still slow, on the other hand, power consumption is high, and performance and capacity are strongly related to the process. The existing memory cannot keep up with the future demand for high read/write speed, low latency, and low power consumption of data.
ReRAM can improve read/write performance by 100 times compared to NAND, while maintaining lower power consumption and high storage density, which is expected to solve the future demand for high energy efficiency and low latency in data centers and achieve higher performance AI data centers.
4.AI computing (storage and computing in one)
Artificial intelligence is an important trend in current technology development. According to Sullivan Consulting data, the average annual growth rate of AI from 2016-2024 reached 33.98%, and is expected to exceed $615.7 billion in 2024.
Arithmetic power, algorithms, and data volume are the three basic elements of AI development, which determine the performance of AI computing, two of which are storage-related: data is carried by memory, and data volume determines the accuracy of AI computing models; arithmetic power, the future of the chip computing performance and latency have put forward higher requirements.
With the current von Neumann architecture, the storage unit and the computing unit are independently separated, the process of moving data consumes a lot of time and energy, and due to the different process routes of the processor and the memory, the data access speed of the memory can hardly keep up with the data processing speed of the CPU, and the performance has lagged far behind the processor. Therefore, the von Neumann architecture has natural limitations in terms of data processing speed and energy efficiency ratio, which is called “storage wall”.
By integrating the storage and computation units into one, the storage and computation architecture eliminates the latency and power consumption caused by data access and can break through the “storage wall” to achieve higher computing power and higher energy efficiency.
Read more:
