(Report Producer/Author: Huaxin Securities, Mao Zheng, Zhang Lu)
存储紧随AI步伐， H B M 应运而生
The share of DRAM ranks first in the storage market
Memory is a part used to store programs and data, for computers, with memory, in order to have memory function, in order to ensure normal operation. There are many types of memory, and according to the different media, memory is divided into three categories: optical storage, semiconductor storage and magnetic storage.
Semiconductor memory can be divided into volatile memory chips (RAM) and non-volatile memory chips (ROMs). Among them, volatile storage is mainly SRAM (static random memory) and DRAM (dynamic random memory). Non-volatile storage ranges from early ROMs (PROM, EPROM, EEPROM) to flash memory (NOR Flash and NAND Flash). According to key parameters such as transmission speed, capacity, and data erasure, different memory has its own applicable fields, in terms of use in computer systems: SRAM is fast enough and is usually used as a cache for the central processing unit (CPU), DRAM is cheap and has a large capacity, and is usually used as memory, NAND Flash is used for big data storage, and NOR Flash is usually used for code storage. In terms of market share of storage industry products, DRAM ranks first with a share of more than 50%, followed by NAND with a share of about 41%. NOR remains stable at about 1% of the market share, while other products such as EEPROM account for about 1% of the market share.
GDDR and HBM are suitable for high-performance computing
Before the advent of HBM, DRAM was mainly classified into DDR, LPDDR, and GDDR according to product classification. Among them, DDR is suitable for computers, servers and other high-performance computing equipment; LPDDR is suitable for long-term use scenarios such as mobile devices and embedded systems, such as mobile phones, tablets, and wearables. GDDR is a high-performance DDR memory specification specially designed for the design of high-end graphics cards, which has a higher clock frequency and less heat than ordinary DDR memory used in the main memory, so it is more suitable for high-end display chips.
In the new demand of the AI era, HBM came into being
With the rapid development of artificial intelligence technology, traditional GDDR memory has gradually reached the bottleneck of its technological development. 1) GDDR5 can't keep up with GPU performance: the number of parameters trained by AI increases by 410 times every two years, while the memory of a single GPU only grows at a rate of 2 times every two years. Hardware has increased its peak computing power by 60,000 times in 20 years, but DRAM bandwidth has only increased by 100 times and interconnect bandwidth has increased by only 30 times. It is difficult for memory to keep up with the computing speed of AI hardware, which limits the performance of AI chips and forms a "memory wall". 2) GDDR5 limits form factor: In order to achieve high bandwidth, more and more GDDR5 and circuits need to be integrated together, which will limit the size of the product; 3) On-chip integration is not a panacea: The compatibility and scalability of on-chip integration are limited. The emergence of HBM overcomes the bandwidth limitations of traditional GDDR and breaks through the limitations of memory walls to meet the demand for high-computing power and high-bandwidth memory in the AI era.
The demand gap at home and abroad has widened, and the volume and price of HBM have risen
HBM achieves high bandwidth, low latency, and low power consumption
HBM stands for High Bandwidth Memory, which stacks multiple DDR chips vertically on top of each other and encapsulates them with the GPU. Compared with traditional DRAM, it ensures more I/O (Input/Output) channels, can transmit a large amount of data at a time, and achieves high bandwidth, low latency, and low power consumption in a small physical space, making it a next-generation memory solution for application scenarios such as data centers.
The three major overseas storage factories are divided into three parts of HBM in the world
The global HBM market has long been firmly dominated by the three major storage OEMs. According to Goldman Sachs, in the competitive landscape of the HBM market in 2023, Hynix accounts for about 54%, Samsung accounts for 41%, and Micron accounts for 5%. Hynix currently dominates the market, Samsung is following suit, and Micron, although currently having a smaller market share, is aggressively expanding its production capacity and technology. Due to the lack of HBM3E production capacity, the three giants may all enter the NVIDIA supply chain, at present, Micron is the first to obtain the Nvidia H200 order, Hynix has also begun mass production and supply to NVIDIA, Samsung has passed the verification of AMD MI350, and the verification time of Nvidia is slightly later than the other two.
At the beginning of HBM in China, the three giants promoted the construction of production lines
Most of the domestic DRAM market share is still occupied by Sanxing, Hynix and Micron, and some Taiwanese manufacturers such as Nanya Technology and Winbond Electronics compete for market share, while the market share of domestic manufacturers is relatively small, and the main DRAM manufacturers are GigaDevice, Beijing Junzheng, Dongxin, Changxin Storage, Unisplendour Guowei, Fujian Jinhua, etc. The localization of HBM in China has just started, and the storage giants are accelerating. Domestic storage manufacturers Wuhan Xinxin and Changxin Storage are in the early stages of HBM manufacturing, mainly to meet the application needs of future artificial intelligence (AI) and high-performance computing (HPC) fields. Among them, Wuhan Xinxin is building a 12-inch factory with a monthly production capacity of 3,000 wafers for HBM, while Changxin Storage has cooperated with Tongfu Microelectronics, a packaging and testing factory, to develop HBM samples and show them to potential customers.
Driven by strong demand for AI, HBM volume and price have risen
Demand in the field of AI and high-performance computing (HPC) is the main driver of HBM's growth. HBM's remarkable high speed, high bandwidth, and low power consumption make it a promising application in the field of AI and high-performance computing. Mainstream GPUs for AI servers, such as H100, A100, and A800, as well as AMD MI250 and MI250X, are basically equipped with HBM. The subsequent ace products have also continued to upgrade the latest iteration of HBM, and the demand for HBM remains high. The world's major technology giants are bidding for HBM, and the HBM market will grow rapidly under the wave of AI. According to Tencent Technology's estimates, in 2024, the global production capacity of GPU products packaged through CoWoS will be about 9 million. According to the standard calculation of a single GPU logic chip with 6 HBM memory chips (some with 8 HBMs), the global GPU in 2024 will have a demand for more than 54 million HBM memory chips (mainly 12 layers). TrendForce expects global HBM demand to grow by nearly 60% annually in 2023 and another 30% in 2024.
The three key processes occupy the value of the industrial chain
TSV: Nearly half of the etching cost
TSV is the core process in HBM's production process. TSV (Through Silicon Via) is a technology that realizes the interconnection of chips by making vertical conduction between chips and wafers. TSV technology allows vertical interconnects to be created inside silicon wafers, which is fundamental to enabling multi-layer DRAM chip stacking. It reduces the interconnection length, signal delay, capacitance and inductance through vertical interconnection, realizes low-power and high-speed communication between chips, increases bandwidth and realizes the miniaturization of device integration. TSV is also the link with the highest proportion of value. Taking the HBM structure of four 5mmx7mm memory chips stacked on a logic chip as an example, assuming that the yield of thin wafer processing and the yield of TSV formation are 98%, the cost of TSV process accounts for about 30% of the overall packaging cost, which is the largest part of the process.
Bump Fabrication: Towards Hybrid Bonding
Bumping is the production of bumps on the chip, which provides a "point" interface for the electrical interconnection of the chip by making metal bumps on the surface of the chip, which is widely used in advanced packaging such as FC, WLP, CSP, and 3D. In this process, the UBM layer is formed as a whole and used as a conductive layer for plating, and then photoresist is used to protect the areas that do not need plating. Once the plating has formed a thick bump, the photoresist is removed and the unwanted UBM is etched, which is then reflowed to form a bump array.
The bump pitch is advanced to less than 10 μm, and hybrid bonding is headed. The smaller the bump spacing, the higher the bump density, the higher the package integration, and the greater the difficulty. The minimum distance between the bumps is 40 μm; When the pitch < 40 μm, a hot press bonding (TCB)-based joining technology is used; However, when moving towards smaller bump pitches (less than 10 μm), hybrid bonding, which forms a direct Cu-Cu bond with fine pitches (<1-20 μm), is the most competitive interconnect method and will widely replace micro-bumps and copper pillar bumps in the future.
堆叠：MR-MUF助力海力士
In the HBM stacking process, it is often necessary to use some materials to protect the chips and interconnects, as well as to ensure the mechanical stability and electrical properties of the stacking structure. Conventional HBM uses TC-NCF (Non-Conductive Film) for stacking, in which a non-conductive adhesive film is placed between the layers each time to isolate the chips from each other and protect the connection points from impact. At present, both Samsung and Micron continue to use TC-NCF, while Hynix uses another MR-MUF (Mass Reflow Molded Underfill), which uses EMC (Epoxy Molding Compound) to fill and paste between chips when they are stacked upwards. Compared to the TC-NCF, MR-MUF reduces the die stacking pressure to 6%, shortens the process time, increases the production efficiency by 4 times, and improves the heat dissipation rate by 45%.
Analysis of key companies
AMEC: a leader in semiconductor equipment, continuing to deploy high-end products
Founded in 2004, AMEC is mainly engaged in the R&D, production and sales of semiconductor equipment, and went public in 2019. The company mainly provides etching equipment, MOCVD equipment and other equipment for the manufacturing of integrated circuits, LED chips, MEMS and other semiconductor products. The company's plasma etching equipment has been applied to international first-line customers from 65nm to 14nm, 7nm and 5nm and other advanced integrated circuit processing and manufacturing production lines and advanced packaging production lines. The company's MOCVD equipment has been put into mass production on the production lines of industry-leading customers, and the company has become the world's leading manufacturer of gallium nitride-based LED equipment.
Tuojing Technology: CVD equipment leader, hybrid bonding equipment to open the second growth pole
Founded in 2010, Tuojing Technology is one of the important leading enterprises in the domestic semiconductor equipment industry, and the company has won the title of "China's Top 5 Semiconductor Equipment Enterprises" issued by the China Semiconductor Industry Association three times (2016, 2017 and 2019). The main products include plasma-enhanced chemical vapor deposition (PECVD) equipment, atomic layer deposition (ALD) equipment and sub-atmospheric chemical vapor deposition (SACVD) equipment, which have been widely used in domestic wafer fabs 14nm and above process integrated circuit manufacturing production lines, and have carried out 10nm and below process product verification tests. THE COMPANY'S PRODUCTS HAVE BEEN ADAPTED TO THE MOST ADVANCED 28/14NM LOGIC CHIPS, 19/17NM DRAM CHIPS AND 64/128-LAYER 3DNANDFLASH WAFER MANUFACTURING LINES.
Saiteng shares: consumer electronics, semiconductors two-wheel drive performance growth
Founded in 2001, the company has been focusing on high-tech enterprises in the field of automation equipment, and has strong competitive advantages and independent innovation capabilities in consumer electronics, semiconductors and other intelligent assembly, testing, measurement, etc., and has a number of independent research and development of core technological achievements. The company has become a high-quality partner of many well-known enterprises at home and abroad. The company's automation equipment mainly includes two categories: non-standardized automation equipment and standardized automation equipment, which are mainly non-standardized automation equipment in the consumer electronics industry; In the semiconductor industry, it is mainly industry standard equipment, and specific products such as die bonding equipment, sorting equipment, wafer packaging machines, wafer defect detection machines, chamfer roughness measurement, wafer character inspection machines, wafer laser marking machines, and wafer laser grooving machines.
Excerpts from the report:
(This article is for informational purposes only and does not represent any investment advice from us.) To use the information, please refer to the original report. ）
Selected report source: [Future Think Tank]. Future Think Tank - Official Website