The inspiration that Japan and South Korea's DRAM technology brings to domestic enterprises

DRAM memory and Flash memory chips are the most important memories in the current market. DRAM is the most common system memory. Although it has excellent performance, it is easy to lose power. Compared with its volatile memory of the same level, its cost is lower, so it is the most common in system memory; Flash is the most widely used non-volatile storage. Its non-volatility when power is off makes it mainly used in the field of large-capacity storage.

DRAM: Common storage medium for memory

DRAM can only store data for a very short time. It uses capacitors to store data, so it must be refreshed at regular intervals, otherwise the information will be lost.

Compared with SRAM, DRAM is slower and retains data for a relatively short time, but its price is cheaper. Due to the difference in technology, DRAM has lower power consumption, higher integration and smaller size, and is also faster than all ROMs, so it is widely used.

Flash: Large capacity flash memory

In the early days of memory development, ROM has always been the main storage device of the system, but it has now been completely replaced by Flash. In terms of characteristics, Flash has the advantages of both RAM and ROM. Not only does it not lose data after power failure, but it also has electronic erasable and programmable properties.

Although Flash is slightly inferior to DRAM in terms of reading speed, it is still fast and its cost is much lower than DRAM.

In terms of classification, Flash is currently divided into two types: NOR and NAND. The main difference between the two lies in the difference in the reading method and the different connection methods of the storage units.

(1) NORFlash:

NORFlash uses "word" as the basic unit and can directly run the code loaded in it (XIP). Compared with NAND Flash, NORFlash has a slower writing speed and higher cost, so it is mainly used in small-capacity code flash memory fields such as DVD, feature phones, USB Key, TV, set-top boxes, and IoT devices. Currently, NORFlash occupies most of the market share in the 0-16MB Flash market.

NORFlash can be divided into serial and parallel. Currently, serial is the main type with XIP features, but the cost is relatively high and it mainly occupies the small capacity market. In addition, due to the simple interface, thinner and smaller size, lower power consumption and overall system cost of serial, although it is not as fast as parallel NORFlash in reading speed, it has become the first choice of major system solution providers.

(2) NAND Flash:

NAND Flash uses "block" as the basic unit, and its unit capacity cost is low, and its writing and reading speeds are better than NORFlash, but users cannot directly run the code on NAND Flash. Therefore, many development boards need another NORFlash to run the startup code while using NAND Flash. Due to the characteristics of NAND Flash such as fast writing and erasing speed and low cost, it is mainly used in the field of large-capacity storage, such as DOC (chip disk) of embedded systems (non-PC systems), and commonly used flash drives, such as mobile phones, tablets, USB flash drives, solid-state hard drives, etc.

DRAM can be divided into DDR series, GPDDR series, LPDDR series and other categories according to different technical specifications. Among them, DDR series is ordinary DRAM, and GPDDR stands for Graphics Double Data Rate Memory (Graphics Double Data Rate), which is a synchronous dynamic random access memory used by high-performance graphics cards and is designed for computer applications with high bandwidth requirements.

LPDDR refers to Low Power Double Data Rate SDRAM, which is mainly used in portable devices. Currently, DDR and DDR2 have basically withdrawn from the market, and DDR3, DDR4 and LPDDR series are the main ones.

DDR3 is a memory product in the SDRAM family. It provides higher operating performance and lower voltage than DDR2. It is the successor of DDR2 (increased to eight times) and is also the most popular memory product specification.

DDR3 uses an 8-bit pre-fetch design, while DDR2 uses a 4-bit pre-fetch design. This means that the DRAM core frequency is only 1/8 of the equivalent data frequency, and the core operating frequency of DDR3-800 is only 100MHz. Secondly, DDR3 uses a point-to-point topology to reduce the burden on the address/command and control buses. Finally, DDR3 uses a production process below 100nm, reducing the operating voltage from DDR2's 1.8V to 1.5V, and adding asynchronous reset and ZQ calibration functions.

DDR4 memory is the latest DDR series memory specification on the market. The first DDR4 memory was successfully developed by Samsung in 2014. There are three major differences between DDR4 and DDR3: 16-bit pre-fetch mechanism (DDR3 is 8-bit), theoretical speed is twice that of DDR3 at the same core frequency; more reliable transmission specification, data reliability is further improved; working voltage is reduced to 1.2V, which is more energy-efficient.

In the future, DDR5 specifications will also be available. In October 2018, Cadence and Micron announced their DDR5 memory research and development progress. The two manufacturers have begun developing 16GB DDR5 products and plan to achieve mass production goals by the end of 2019.

The main feature of DDR5 is chip capacity, not just higher performance and lower power consumption. DDR5 is expected to bring I/O speeds of 4266 to 6400MT/s, with a power supply voltage down to 1.1V. Compared to DDR4, the improved DDR5 features will increase the actual bandwidth by 36%, and even starting at 3200MT/s and 4800MT/s speeds, the actual bandwidth will be 87% higher compared to DDR4-3200. At the same time, one of the most important features of DDR5 will be a single-chip chip density of more than 16GB.

LPDDR (Low Power Double Data Rate SDRAM) is a type of DDR SDRAM, also known as mDDR (Mobile DDR SDRAM). It is a communication standard developed by the American JEDEC Solid State Technology Association for low-power memory. It is known for its low power consumption and small size and is specifically used in mobile electronic products. DDR/DDR3/DDR4/DDR5 are memory particles, and memory sticks are made by embedding multiple particles together in a board, which are used in computers, etc.

Market volatility weakens and demand strengthens, and 2020 may return to balance

DRAM prices have entered a downward cycle since the second half of 2018. Through supply and demand analysis, we believe that DRAM supply is currently in a low-growth plateau due to the slowdown in technology node progress, while demand will maintain a high growth rate driven by applications such as 5G, AI, and big data. Therefore, we expect the DRAM market to digest inventory in 2019 and reach a balance between supply and demand around 2020.

DRAM remains popular and occupies half of the storage market

DRAM is an evergreen in the memory market. Since IBM developed the world's first volatile memory (DRAM) in 1966, it has always occupied a core position in our computing systems. From the perspective of the existing computer system structure, memory is divided into three categories: cache, main memory, and external memory (auxiliary memory). Among them, cache requires high speed but small capacity, and SRAM is usually used.

Memory requires a certain read and write speed and space to support the running program itself and the required data. Compared with SRAM, DRAM retains data for a shorter time and is relatively slow, but in terms of price, DRAM is much cheaper than SRAM. Due to technical differences, DRAM is small in size, highly integrated, and has low power consumption. At the same time, its speed is faster than all ROMs, so it has always been the only choice for memory.

As for external memory, it is equivalent to the computer's data warehouse. It does not require as fast reading and writing speed as the first two, but has a huge demand for capacity. Among the three major memory media, DRAM has the most stable position and the largest market. Although SRAM is expensive, its capacity has increased very little over the years, and it only needs to meet the data transmission within the computer. The demand for external memory capacity is growing too fast, resulting in the need to constantly look for new media.

With the massive generation of data and the trend of miniaturization of electronic devices, the media of external storage has been changing to meet the needs, from magnetic disks/optical disks/hard disks to Flash and SSDs. Only DRAM has the characteristics of high density and high capacity since its birth. From the initial K level to the current GB level, the principle of DRAM itself has not changed much.

In the era when semi-magnetic storage media were used as external memory, DRAM was almost synonymous with semiconductor memory. After entering the new century, the development of portable devices and the maturity of semiconductor technology pushed the memory competition to evolve into a two-pronged battle between DRAM and Flash. In this process, it can be divided into the NOR era and the NAND era.

Before the popularity of smart phones, portable devices did not require much storage space. In addition, NOR Flash's feature of supporting random access allowed it to execute programs like ordinary ROM, making it the mainstream storage medium for portable devices. In 2002, DRAM accounted for 55% of the sales of the entire memory market, while NOR Flash accounted for 21%. NAND Flash only accounted for 8%, and was mainly used in applications that require larger storage space, such as MP3, SD cards and USB flash drives.

External memory is ever-changing, but DRAM is always there. In the era of smart phones and portable devices, the situation has changed fundamentally, and NAND Flash has quickly replaced NOR Flash to become the mainstream of flash memory.

From 2008 to 2018, the continuous increase in smartphone shipments and the continuous expansion of single-machine storage capacity have become one of the main forces driving the continuous expansion of DRAM and NAND Flash demand.

According to Yole estimates, DRAM accounted for 61% of the memory market in 2018, while NAND Flash accounted for 36%. Only 5% was left for NOR Flash, ROM, and SRAM. In the process of external memory media reshuffle, DRAM's market share has remained above 50%, fully reflecting its technical scalability and huge market demand.

Diversified demand leads to weaker cyclical fluctuations, and it is expected to return to balance in 2020

From the perspective of market performance, the memory market shows obvious cyclical fluctuations. However, from the perspective of long-term trends, the amplitude of cyclical fluctuations is gradually decreasing, and the overall upward trend of the industry is clear.

Looking at the memory market over the past 30 years, the memory market has experienced four major corrections, namely between 1995 and 1998, the bursting of the Internet bubble in 2001, between 2007 and 2008, and between 2011 and 2012. The magnitudes of these four corrections were 58%, 50%, 23%, and 17%, respectively. The main reason behind this is the diversification of the demand side structure and the faster rebalancing of supply and demand.

From the demand side, the demand structure of memory is rapidly changing towards diversification. The demand for server DRAM brought by applications such as cloud services and big data will become a strong growth driver for the DRAM market in the future. Currently, DRAM applications are divided into mobile devices, servers, PCs, consumer electronics and other major fields. According to DRAMeXchange statistics, DRAM demand increased by 22.3% in 2018, among which server applications maintained the fastest growth rate for two consecutive years.

The growth pattern of smartphone DRAM demand has changed from the "double increase" of shipments and single-unit capacity to the single-point pull of capacity increase. In 2018, the DRAM bit demand will change from the single-point pull of smartphone demand in the past to the simultaneous growth of smartphone demand and server demand.

Against the backdrop of sluggish growth in smartphone shipments, the growth in smartphone memory capacity has become the main driver of the growth in mobile DRAM demand. Although mobile DRAM will remain the most important DRAM market in the future, its growth rate will slow down due to changes in growth patterns.

Server DRAM demand rises rapidly

In 2018, global server shipments were approximately 12.42 million units, a year-on-year increase of 5%. Although the shipment growth rate seems low, the capacity of a single machine is rising rapidly. According to DRAMeXrange, the average memory load of servers reached 145GB in 2018, and it is expected that the average DRAM capacity of standard servers will reach 366GB by 2021, with a CAGR of 26%.

Another trend in the server field that cannot be ignored is the rapid development of data centers. Compared with the overall growth rate of servers below 10%, the growth of data centers is as high as 20%. According to DRAMeXchange statistics, an average IDC can accommodate about 8,000 to 15,000 server racks, and a rack can carry more than 4 servers of different sizes. It is estimated that this will drive the demand for server DRAM bits of 10 million GB to 2 million GB.

In addition to traditional servers, servers with special requirements such as deep learning will also strongly drive storage demand. According to Micron's estimates, a server used for AI training requires 6 times more DRAM and 2 times more SSD than an ordinary server.

In 2021, the shipment of servers with AI training capabilities will reach one-eighth of the global server shipments, and this proportion is expected to triple by 2025. With the combined effect of the above factors, the demand for DRAM and NAND Flash will continue to maintain a compound annual growth rate of 20% and 40% in the next two years.

From the supply side, the growth rate of DRAM supply is in an overall slowing trend. The growth source of DRAM bit supply is mainly due to the increase in density brought about by process advancement, supplemented by the increase in wafer input brought about by capacity expansion.

However, in recent years, after DRAM entered the 20nm process, the process upgrade began to encounter bottlenecks. Mainstream manufacturers are cautious about the development and application of 1Xnm and below processes due to cost and R&D difficulty. Currently, Samsung, Micron, and Hynix are struggling to advance from 20nm to 18nm. Taiwanese manufacturers, except Nanya Technology, still mainly use 38nm process. The slowdown in process advancement and the reduction in storage density growth have directly led to a decline in the growth rate of DRAM comprehensive bit supply.

DRAM process advancement slowed down, and production capacity fluctuations remained basically stable. Global DRAM production capacity and wafer input experienced a significant reshuffle between 2010 and 2013.

In 2010, 40nm DRAM products began to enter the mainstream market, and in the following three years, the process technology frontier quickly advanced to 20nm. Samsung and Hynix, which led the technology transition, gained the dual advantages of storage density and cost while maintaining production capacity, causing other manufacturers to lose market share. Elpida, the fourth largest DRAM manufacturer at the time, went bankrupt and was acquired by Micron.

From the supply side, 2013-2017 was a period of plateau production capacity, with overall production capacity stable and the proportion of 20nm process gradually increasing. DRAM prices fell first and then rose during this period, mainly due to the absorption of the abundant supply brought by the previous process improvement. The fundamental difference between the current weakness of the DRAM market and that in 2013 is that there is no leapfrog development of the process and the supply and demand relationship has not changed qualitatively.

As the bit supply growth rate slows down, it will gap with the demand growth rate. We expect that the previous inventory and slight oversupply will be digested around 2020 and the balance will be restored.

After 2020, the popularization and application of 5G and AI will become an important force driving the demand for semiconductors. At the same time, the next-generation DRAM process will also begin to become popular. The supply and demand relationship of the entire DRAM market will become more complicated, but the overall upward trend in scale is certain.

Except for Hynix's new Wuxi production line, other manufacturers have no large-scale expansion plans from 2019 to 2020, and the overall annual wafer output growth rate is between 3% and 5%. On this basis, we comprehensively consider the impact of previous years such as supply and demand growth and accumulated inventory to calculate the supply and demand balance.

In 2016, DRAM prices turned from falling to rising, so 2016 is taken as the year of supply and demand balance, and the supply and demand indexes are both 100, and the annual supply index already includes the inventory situation of previous years. In 2016, major manufacturers basically completed the 20nm process conversion, ending the technology-led supply growth from 2013 to 2016. As a result, the overall bit supply and demand growth rate declined in 2017, resulting in a supply gap.

In 2018, Samsung expanded production by 8%, and Hynix's Wuxi plant also expanded production slightly, quickly filling the demand gap and ending the boom. However, after that, except for Hynix, other major manufacturers did not expand production on a large scale, and processes below 1Znm are expected to enter the market on a large scale in 2021. This year and next year will be a dual platform period for wafer input and process level. We expect that the demand growth rate will outpace the inventory in 2019, and the supply and demand of DRAM bits will be balanced again around 2020.

It should be noted that because the demand for DRAM bits is rigid, even a small imbalance between supply and demand will be magnified by market behavior into a large price fluctuation. The supply gap or excess demand we calculated is basically below 5% each year, but the direct reflection of prices in the market fluctuates violently. The rush to order and stock up during price increases not only pushes up prices, but also accumulates inventory. These inventories will rush to clear out when demand growth slows down slightly and a price inflection point is formed, exacerbating the price decline.

DRAM market applications are constantly being innovated, and the downward cycle is generally controllable, and it still has strong vitality and value. The question now is how Chinese storage DRAM companies should face such a huge market that is monopolized by overseas giants? We believe that the barriers to the memory industry are high, but they are not easily broken.

Judging from the rise of storage giants in history, the development path of technology introduction + independent research and development of industry, academia and research + comprehensive support is feasible. In addition, China has a huge demand market, which makes it easy to form a virtuous closed loop of the industry, which is also an important advantage that other countries do not have.

Exploring China's storage development path from the development history of the storage industry

As a technology-intensive, capital-intensive, and highly monopolized industry, the memory industry has always been unfriendly to latecomers. For Chinese memory companies to grow and develop, in addition to market demand feasibility, they also need to invest a lot of resources in technology research and development and be prepared for higher-level strategic games. This is not only China's path, but also the path of every latecomer in the history of the memory industry.

Looking back at the 50-year development history of the semiconductor memory industry, it can be roughly divided into three periods: the US-dominated period from 1970 to 1982; the Japanese-dominated period from 1982 to 1998; and the South Korean-dominated period from 1998 to the present.

With the exception of the United States, the rise of the storage industry in the other two countries is deeply bound to various social forces and overall economic development. The development of the storage industry has also evolved from simple "primary technology drive" to "government, industry and academia joint technology drive" and gradually to "government, industry and academia joint technology drive + multi-faceted long-term support".

The era of American dominance: the dawn of semiconductor storage driven by original technology

Unlike Japan and South Korea, when the United States developed memory, personal computers were not yet popular. Therefore, the amount of memory used was small and the price was high. The development of memory was far from commercial wars and was more driven by technology.

In 1969, with the efforts of the first generation of integrated circuit pioneers such as Noyce and Moore, Intel successfully developed the first memory chip - the 3101 chip with a capacity of 64 bytes.

The following year, Intel's 12th employee, Ted Hoff, proposed a new design that reduced the number of transistors in a DRAM memory cell from four to three. This allowed more memory cells to be packed together, greatly increasing the storage space to 1024 bytes. This was the technical prototype of the DRAM we use today.

In 1970, Intel made further progress in memory research and development, and they developed an erasable programmable read-only memory (EPROM) with a capacity of 2K. In 1972, Intel further developed the world's first static random access memory (SRAM) 2102 chip. In the 1970s and 1980s, the capacity of memory increased exponentially, and 4K, 16K, and 64K DRAM chips were successively launched. Semiconductor memory during this period was basically monopolized by American companies such as Intel and MOSTEK.

The rise of Japanese storage: creating a development model of "government, industry and academia"

As a latecomer, Japan pioneered the top-level design to protect the semiconductor industry. In the 1970s, the Japanese government promoted the development of domestic semiconductor strength through the integration of industry, government and academia, and protected the industry through import barriers.

Japan's semiconductor storage industry started relatively early. In 1971, NEC launched DRAM chips, closely following Intel's mass-produced DRAM. Despite this, Japan's semiconductor technology and product performance still lag far behind those of the United States. At the same time, the United States already used very large-scale integrated circuits (VLSI) for memory, while Japan was still stuck in the previous generation of large-scale integrated circuits (LSI).

In 1976, the Japanese government's Ministry of International Trade and Industry took the lead, with five major companies, Hitachi, Mitsubishi, Fujitsu, Toshiba, and NEC, as the backbone, and joined the Electrical Engineering Laboratory (EIL) of the Ministry of International Trade and Industry of Japan, the Electronics Research Institute of the Japan Industrial Technology Research Institute, and the Computer Research Institute, and invested 72 billion yen to tackle the technical difficulties of very large-scale integrated circuits DRAM. The four-year VLSI project achieved remarkable results. Teams from different companies shared their knowledge and competed with each other, and obtained a total of 1,210 patents and 347 trade secrets.

The success of the VLSI research project directly helped Japan surpass the United States in terms of DRAM cost and reliability. In the late 1970s, the yield rate of DRAM in the United States was around 50%, while Japan was able to achieve an astonishing 80% at the time, which constituted an overwhelming overall cost advantage. As a result, Japanese storage companies took advantage of the victory to provoke a price war. The price of DRAM chips dropped from US$50 in 1981 to US$5 per chip in 1982. American manufacturers could not withstand it and retreated step by step. It was during this period that the Japanese storage industry completed its reversal of the United States. In the heyday of the late 1980s and early 1990s, Japanese DRAM accounted for more than 65% of the global DRAM market share, eventually forcing Intel out of the DRAM market.

The most valuable experience we have gained from the rise of Japan's memory industry is that it reveals the technology-intensive and capital-intensive characteristics of the memory industry, and demonstrates the feasibility and importance of the joint development of the memory industry by government, industry and academia.

After the idyllic era of memory in the 1960s, the memory market has grown rapidly and the technical barriers have risen rapidly. In the subsequent competition, the requirements for the three major factors of technology, capital, and market are extremely strict. It is difficult for a single enterprise to catch up, so latecomers must succeed through the joint efforts of enterprises and governments.

The rise of Korean storage: R&D + support to win the protracted war

The early development of South Korea's semiconductor industry relied on low labor and land costs to attract foreign investment and build factories. During this period, South Korea quickly accumulated a large amount of capital and formed the embryonic form of the semiconductor industry. However, the low-end development model that lacked technology and was labor-intensive came to an end in the 1970s.

In order to promote industrial upgrading, the South Korean government announced the "Heavy Industry Promotion Plan" in 1973 and announced a six-year plan to support the semiconductor industry in 1975, emphasizing the localization of electronic components and semiconductor production.

With Japan's successful experience, South Korea knows that it must master core technology to go a long way in the storage industry. During the "Semiconductor Industry Revitalization Plan" from 1982 to 1987, South Korea followed Japan's VLSI breakthrough project, led by the Korea Electronics and Telecommunications Research Institute, in conjunction with Samsung, LG, Hyundai and six Korean universities, to tackle DRAM technology. The project lasted three years and the R&D cost reached US$110 million, of which the South Korean government borne 57%.

In addition to catching up with technology, the establishment of South Korea's memory hegemony is also inseparable from historical opportunities and brutal commercial competition. The biggest historical opportunity that the Korean memory industry has seized was the US-Japan semiconductor dispute in 1987. The dispute ended with Japan giving in and promising to increase chip prices by reducing DRAM production. However, at this time, the wave of computer popularization coincided with the reduction of DRAM production, which caused a huge gap in the global 256KDRAM. Korean memory companies seized the opportunity and filled the market gap.

In terms of business war, South Korea's determination and strength can be described as desperate, not giving up until the goal is achieved, and ignoring the long-term huge losses, it is determined to stick to the storage industry. For example, when Samsung launched 64KDRAM in 1984, it coincided with the global semiconductor industry downturn. The price of memory fell from $4 per piece to 30 cents per piece, while Samsung's production cost at the time was $1.3 per piece, which means that Samsung lost $1 for every memory sold.

Samsung continued to suffer huge losses for nine consecutive years in the 1990s, and its debt ratio once reached 300% during the Asian financial crisis. During this period, the Korean government and domestic financial groups supported Samsung with financial strength. The Korean government alone provided more than US$6 billion in policy loans at preferential interest rates.

The market and production capacity supply are all in China, and the rise of the domestic memory industry is an inevitable trend. On the supply side, mainland China has become the region with the largest scale and fastest growth in semiconductor capital expenditure in the world. In the next few years, mainland China's wafer capital expenditure will be in the hundreds of billions, and the proportion of domestic capital will increase significantly, especially memory factories represented by Hefei Changxin and others will be put into production one after another. On the demand side, China's semiconductor consumption has accounted for more than 60% of the world, and the market share of domestic smartphones, PCs, and servers has continued to increase. In the future, it is expected to quickly absorb domestic memory production capacity and form a virtuous closed loop.