1. Memory chips are one of the integrated circuit products with a huge market size.
(1) Memory chips belong to the category of integrated circuits in semiconductors and are one of the most widely used and standardized basic products of integrated circuits. Semiconductors can be classified into four categories according to product classification: optoelectronic devices, sensor devices, discrete devices, and integrated circuits. Integrated circuits account for the largest proportion of the value of semiconductors, accounting for about 82.64% of the entire semiconductor industry market size. They mainly include four types: analog chips, microprocessor chips, logic chips, and memory chips. According to WSTS data, the global semiconductor market size in 2022 was US$574.084 billion, with integrated circuits accounting for 83%. Among them, the market size of memory chips was US$129.767 billion, accounting for 23% of the entire semiconductor industry. It can be seen that memory chips and logic chips contribute the largest value in the entire semiconductor industry chain.
Semiconductor product classification diagram
Data source: WSTS, compiled by Donghai Securities Research Institute
(2) Storage devices are hardware components in a computer system used to store and retrieve data. They can be classified into optical storage, magnetic storage, and semiconductor storage based on the storage medium. Optical storage refers to a device that uses optical methods to read and store data from an optical storage medium, generally referring to optical disc drives, optical tape drives, and optical card drives. Magnetic storage refers to magnetic media devices that use magnetic energy to store information; its storage and retrieval process requires the mechanical movement of magnetic disks, and it is currently widely used in PC hard drives, portable hard drives, and other fields. Memory chips, also known as semiconductor memory, refer to semiconductor media devices that use electrical energy to store information; their storage and retrieval process involves the storage or release of electrons, and they are widely used in memory, USB flash drives, consumer electronics, smart terminals, solid-state storage hard drives, and other fields. Based on whether the stored information is retained after power failure, memory chips can be mainly divided into volatile memory chips (RAM) and non-volatile memory chips (ROM).
RAM stands for Random Access Memory, which does not retain data after power is turned off. Main products include SRAM and DRAM. DRAM, or Dynamic Random Access Memory, uses capacitors for storage; one bit in DRAM is stored using one capacitor and one transistor. Because capacitors can leak current, the memory cell needs to be refreshed periodically to retain data. SRAM, or Static Random Access Memory, has a more complex internal structure than DRAM and can retain data without a refresh circuit. ROM is a type of memory that stores fixed information. Under normal operating conditions, data can only be read and cannot be modified or rewritten in real time. It retains data even after external power is cut off. It has a slower read speed but a larger storage capacity. Main types include EEPROM (Electrically Erasable Programmable Read-Only Memory), Flash memory, PROM (Programmable Read-Only Memory), and EPROM (Erasable Programmable Read-Only Memory).
(3) Based on specific functions, the memory in a computer can be subdivided into six layers: registers, cache, main memory, disk cache, fixed disk, and removable storage media. From CPU cache and main memory to SSD and HDD, the computer's storage system is formed. Each layer only exchanges data with the adjacent layer. As the layer decreases, the device capacity increases, the distance from the CPU increases, the access speed decreases, the transmission time increases, and the cost per byte decreases. The registers in the CPU are located at the top and are denoted as L0. They are made using SRAM chips and integrated inside the CPU. They have limited capacity, extremely fast speed, and are synchronized with the CPU. The cache is a small and fast memory, generally used as a buffer for DRAM. It is implemented using SRAM technology and is usually integrated inside the CPU. Main memory is generally composed of DRAM. Unlike SRAM, it has higher storage density, larger capacity, lower price, and slower speed. In summary, SRAM is expensive and fast, DRAM is cheap and has larger capacity, and SSD and HDD are external storage devices with larger capacity, lower cost, further away from the CPU, and slower access speed.
Memory structure hierarchy diagram
Figure: Storage System Architecture
Source: CSDN, Donghai Securities Research Institute
(4) DRAM and FLASH are currently the most important memory chips on the market. FLASH can be divided into NOR and NAND. The difference between the two lies in the different connection methods of the storage cells, which leads to different reading methods. Because NAND uses pin multiplexing, the reading speed is slightly slower than NOR, but the erasure and writing speeds are much faster than NOR. NAND has a simpler internal circuit, so it has a high data density, small size, and low cost. Some high-capacity FLASH on the market uses NAND, such as SSD, USB flash drive, SD card, and eMMC. Small-capacity FLASH of 2-12M is mostly NOR. NOR is more suitable for frequent random read and write applications and is usually used to store program code and run it directly in the flash memory. Compared with Flash and NOR, DRAM has advantages such as higher read and write speed and shorter storage time, but the unit cost is higher. It is mainly used in PC memory (such as DDR), mobile phone memory (such as LPDDR), and server devices.
Table 3: Differences between Storage Products
Data source: Dongxin Technology's prospectus, publicly available information, Donghai Securities Research Institute
2. Vertical specialization and mergers and acquisitions accelerate supply chain integration.
(1) Memory chips are an important branch of the semiconductor industry with broad downstream application prospects. The upstream participants in the memory chip industry chain include semiconductor material suppliers such as silicon wafers, photoresists, sputtering targets, polishing materials, and electronic specialty gases, as well as semiconductor equipment suppliers such as lithography machines, PVD, CVD, etching equipment, cleaning equipment, and packaging and testing equipment; the midstream of the industry chain consists of memory chip manufacturers, who are mainly responsible for the design, manufacturing, and packaging and testing of memory chips. Common memory chips include DRAM, NAND flash memory chips, and NOR flash memory chips; the downstream of the industry chain consists of companies in application fields such as consumer electronics, automotive electronics, information and communication, and artificial intelligence. Various electronic and intelligent devices cannot function without memory chip applications.
Memory chip industry chain diagram
Data source: China Business Industry Research Institute, compiled by Donghai Securities Research Institute
(2) The memory chip industry chain can be further subdivided into a complete industrial chain according to the manufacturing process. The memory chip industry chain mainly consists of integrated circuit design, wafer manufacturing, packaging and testing, and module manufacturer integration. From the perspective of business model, it is mainly divided into IDM and vertical specialization model. The IDM model refers to the enterprise's business covering all aspects of IC design, manufacturing, packaging and testing. Most of the major international memory chip manufacturers are IDM models, such as Toshiba Semiconductor, Samsung Semiconductor, Spansion Semiconductor, Micron Technology and other large multinational companies. Another model is the vertical specialization model, namely Fabless (design companies without wafer manufacturing) + Foundry (wafer foundry) + OSAT (packaging and testing companies). The Fabless model refers to the integrated circuit design model without a wafer production line. That is, the company only designs and sells integrated circuits, outsourcing the production links such as manufacturing, packaging and testing. Examples include Qualcomm, MediaTek, AMD, and Huada Semiconductor. Foundry is an industry operation model that is only responsible for chip manufacturing and not chip design. Typical representatives of this type of company are TSMC, UMC, and SMIC. OSAT refers to companies that specialize in semiconductor packaging and testing, such as ASE, Amkor, JCET, and Tongfu Microelectronics.
Figure: Complete Industry Chain Diagram of Memory Chip Manufacturing
Data source: China Business Industry Research Institute, compiled by Donghai Securities Research Institute
(3) The vertical division of labor model has been further deepened, reducing costs while significantly improving the efficiency of industrial operations. Under the IDM model, companies invest heavily in the construction of wafer manufacturing equipment and production lines, with collaborative production across all internal links, resulting in significant overall economies of scale and increased gross profit margins. However, under the IDM model, internal management costs increase, making the IDM model suitable for large-scale enterprises. The Fabless model focuses on asset-light operations, is more flexible, and can fully utilize semiconductor industry chain resources, concentrating its main efforts on product R&D and technology iteration, adapting to the fierce market competition environment, and developing rapidly. The disadvantage is that it cannot achieve process synergy and needs to bear various market risks, making it relatively suitable for small businesses. Memory chips have a high degree of standardization, and most international giants operate under the IDM model. Domestic memory companies, due to their smaller scale, initially entered niche markets and mostly adopted the Fabless model. As their scale expands, they may eventually transition to the IDM model. In recent years, large-scale domestic memory projects such as Yangtze Memory Technologies, Hefei Changxin, and Fujian Jinhua are all large-scale wafer fabs operating under the IDM model.
Table 1: Advantages and disadvantages of three industrial chain business models
Source: CSDN, Donghai Securities Research Institute
(4) Industry development and technological upgrading drive changes in the industrial chain model, and mergers and acquisitions accelerate the integration of the industrial chain. Before the establishment of TSMC, there was only one model in the semiconductor industry: IDM. After more than half a century of development, the global semiconductor industrial chain has gradually moved towards the trend of specialization and integration.
1) Industry Chain Division of Labor: Moore's Law drives continuous technological updates and iterations in the semiconductor industry, while also stimulating the upgrading and transformation of production and manufacturing equipment. Advanced process wafer fabs require increased capital investment. The foundry model maximizes the amortization of technology upgrade costs and capacity utilization, improving the return on capital expenditure. To reduce investment risks and achieve asset-light operations, IDM companies are gradually adopting a fabite strategy, outsourcing some non-core and high-difficulty process manufacturing operations to third-party foundries while retaining the remaining manufacturing operations themselves.
2) Supply Chain Integration: The semiconductor industry has entered a new stage of development. Through mergers and acquisitions, companies can expand their product lines and customer base based on business synergies and complementarities, reduce costs, achieve stronger supply chain premiums, and increase industry market share. Therefore, with technological advancements, global division of labor is becoming increasingly prevalent. At the same time, large companies continue to grow and merge into IDM (Integrated Device Manufacturer) models, creating a cyclical industry that drives continuous global technological progress.
Figure 1. Evolution of the Global Semiconductor Supply Chain Model from 1950 to 2010
Data source: TNO and CWTS, Donghai Securities Research Institute
(5) In recent years, with the continuous upgrading of technical difficulty, the share of fabless companies in global IC sales has continued to increase. According to IC Insights data, in terms of sales growth rate, over the past 20 years, the IC sales growth rate of companies adopting the fabless model has been significantly higher than that of companies adopting the IDM model. In some years, such as 2017 and 2018, the share growth of fabless companies was lower than that of IDM companies. Moreover, compared with IDM companies, fabless companies are less affected by market fluctuations. In terms of sales share, from 2003 to 2021, the sales share of global fabless companies in IC sales has steadily increased. In 2019, the share of fabless companies in the IC market increased by 4.1 percentage points year-on-year to 29.9%, and then continued to grow, reaching a peak of 34.8% in 2021. The asset-light nature of the fabless model is more flexible, and it can better adapt to the fierce market competition environment during market downturns. my country has huge potential for domestic production, and most memory chip companies were initially established using the fabless model.
Figure: Sales growth rate (%) of Fabless/IDM companies, 2003-2021
Source: IC Insights, Donghai Securities Research Institute
Figure 2003-2021: Percentage of Global Fabless Company Sales in Total Integrated Circuit Sales (%)
Source: IC Insights, Donghai Securities Research Institute
3. The storage industry has undergone two spatial migrations.
(1) The memory chip industry originated in the United States and has since undergone two major industrial transfers. In the 1960s, with the development of computer technology, the electronics industry began to try to apply integrated circuit technology to the field of computer storage. The memory chip industry chain became increasingly clear from vertical integration to vertical division of labor, and underwent two spatial industrial migrations, from the United States to Japan and then to South Korea.
1) United States: The leading manufacturers in the industry have also changed significantly with the industry migration. Initially, Advanced Memory Systems, a company in California, USA, produced the world's first DRAM chip (with a capacity of only 1KB). Subsequently, memory manufacturers such as Intel, Texas Instruments (TI), Mostek, and Micron gradually grew and developed.
2) Japan: In 1976, in order to develop the semiconductor technology field, Japan established a VLSI joint research and development body through a national system. Subsequently, it successfully developed 64K DRAM, catching up with the United States in research and development. By the 1980s, Japanese manufacturers began to overtake the United States with their advantages in quality and price, and their market share gradually reached the number one in the world. In 1985, as the Cold War between the United States and the Soviet Union continued to weaken, trade friction between Japan and the United States continued to increase. The United States began to suppress the Japanese economy. Under the pressure, Japan's market share in memory chips plummeted and it quickly lost its dominant position.
3) South Korea: South Korean companies seized the opportunity presented by the competition between the US and Japan in the semiconductor industry. With the support of US technology transfer and market access, Samsung Electronics stood out and gradually overtook Japan.
(2) The memory market is developing rapidly, and domestic manufacturers are emerging rapidly. Before 2016, China had no ability to produce DRAM and Flash memory until Hefei Changxin and Yangtze Memory Technologies Co., Ltd. were established. In 2019, Chinese mainland companies began to fully enter the memory market. Yangtze Memory Technologies Co., Ltd.'s 64-layer 3D NAND Flash entered the mass production stage. Subsequently, Hefei Changxin announced the commissioning of China's first 12-inch DRAM factory and announced the first 8Gb DDR4 manufactured by 19-nanometer process. In three years, China has successively overcome the challenges of 3D NAND Flash and DRAM technologies, which has broken the international oligopoly in memory and flash memory manufacturing to a certain extent. China's memory chip industry started late, and it is still a long way to go to achieve global leadership. Advanced manufacturing technology is still in the hands of international giants. Therefore, the development of the memory chip industry requires long-term capital investment and technological innovation to solve the problems of reducing defect rates and increasing production capacity in the production process, and improving performance indicators.
Figure: Development History of Domestic Memory Chips
Data source: China Business Industry Research Institute, Donghai Securities Research Institute
4. Development Trends of Memory Chip Technology
(1) DDR, LPDDR, and GDDR are three memory specifications or standards based on DRAM. The Solid State Technology Association (JEDEC) has defined three DRAM standard categories to help designers meet the power consumption, performance, and specification requirements of target applications. Standard DDR: Targeting server, cloud computing, networking, laptop, desktop, and consumer applications, it supports wider channel widths, higher densities, and different form factors. Its development path is to improve performance by increasing core frequency. Mobile DDR (LPDDR): Targeting mobile electronics and automotive applications, which are highly sensitive to specifications and power consumption, it provides narrower channel widths and multiple low-power operating states. Before the fourth generation, it was based on the development of the same generation of DDR. After the fourth generation, it is based on application-specific development, improving performance by increasing the number of prefetch bits. Graphics DDR (GDDR): Targeting data-intensive applications that require extremely high throughput, such as graphics-related applications, data center acceleration, and AI, it is a high-performance DDR memory used in high-end graphics cards, focusing on data bandwidth and operating frequencies far exceeding those of contemporary DDR.
Figure shows that JEDEC defines three types of DRAM standards.
Figure: Speed comparison and application scenarios of three mainstream memory technologies
Data source: CSDN, Donghai Securities Research Institute, Electronic Enthusiasts, Donghai Securities Research Institute
(2) Due to its performance and cost advantages, DDR has become the mainstream memory for PCs and servers. SDRAM can also be called SDR SDRAM, which stands for Single Data Rate. The core, I/O, and equivalent frequency of SDR SDRAM are the same. It can only be read and written once in one cycle. If it is necessary to write and read at the same time, it must wait until the previous instruction is completed before it can be accessed. DDR is a RAM technology used in the system. Its standard name is DDR SDRAM, which stands for Double Data Rate Synchronous Dynamic Random Memory. Its characteristics are high bandwidth and low latency. Each channel of the DDR bus is 64 bits wide. Each Data pin can be read or written (not simultaneously). The currently launched DDR1-DDR5 are product standards formulated by JEDEC. From the evolution path of DDR1 to DDR5, the power consumption of DDR is getting lower and lower, the transmission speed is getting faster and faster, and the storage capacity is getting larger and larger. The latest standard is DDR5, with a starting rate of 4800MT/S and a maximum of 6400MT/S. The voltage has been reduced from 1.2V to 1.1V, and the power consumption has been reduced by 30%. Samsung has already begun early development of its next-generation DDR6 memory, with its DDR6 design expected to be completed by 2024 and commercial applications to begin after 2025.
3) Benefiting from the rapid development of end-user demand, LPDDR and GDDR have entered a period of rapid iteration. LPDDR, or Low Power Double Data Rate Synchronous Dynamic Random Access Memory, is a type of DDR SDRAM, also known as mDDR (Mobile DDR SDRAM), which has lower power consumption and smaller size than DDR memory of the same generation. The LPDDR5 memory standard commonly used in smartphones was officially released by the JEDEC association in February 2019. Compared with the previous generation LPDDR4 standard, LPDDR5's I/O speed has been increased to 6400 MT/s, which is twice the speed of LPDDR4 and 50% faster than LPDDR4X's 4266 MT/s.
LPDDR Development
GDDR is a RAM technology used for displays, characterized by high bandwidth and high latency. The latest standard is GDDR7. In July 2023, Samsung launched the first GDDR7 graphics memory with a processing speed of 32Gbps.
