Domestic GPUs are on the eve of great changes

1. CPU: central processing unit

CPU (Central Processing Unit) is the core component of the computer hardware system. It is responsible for executing instructions in computer programs, processing and managing various data and tasks in the computer system.

2. GPU: full-featured graphics processor

GPU (Graphics Processing Unit) is called a display chip. It is mostly used in microprocessors that specialize in graphics operations on personal computers, workstations, game consoles, and mobile devices (smartphones, tablets, VR devices). It is the core of computer graphics cards. GPUs are divided into independent and integrated categories. Independent GPUs are generally packaged on independent graphics circuit boards. The production of integrated GPUs with independent video memory is completed by CPU manufacturers, so they have strong compatibility, low power consumption and low heat generation.

The threshold for GPU graphics rendering function is very high. Not only is the hardware complexity very high (much greater than the pure GPGPU function), but the graphics rendering function itself is also fully protected by the intellectual property rights of international giants.

3. GPGPU: parallel computing processor

GPGPU (General-purpose computing on Graphics Processing Unit) is a concept derived from GPU, which refers to a parallel computing processor. It is a subset of full-featured GPU and is only used for general-purpose computing. In addition to graphics processing functions, full-featured GPUs also come with the general computing functions of GPGPUs. If GPGPU manufacturers want to migrate to GPU, there are higher technical barriers and difficulties. At least they need to complete the complete graphics rendering function and some necessary additional functions. Therefore, it is difficult for GPGPU companies to bypass the current technical barriers if they want to complement the graphics rendering function and build a full-featured GPU.

GPU application-side classification : Currently, GPUs are widely used in many industries, including computer games, virtual reality, artificial intelligence, image rendering, scientific computing, etc. According to the application side, GPU can be divided into mobile side and desktop side, and the desktop side is divided into server GPU and PC side GPU.

Server GPUs can be used for professional graphics processing, computing acceleration, deep learning and other applications. According to the development of a series of technologies such as cloud computing and artificial intelligence, server GPUs are mainly independent GPUs.

PC GPU will give priority to integrated GPU in light office scenarios to improve performance; in heavy office scenarios, independent GPU will be given priority to ensure performance output.

Mobile GPUs generally use integrated GPUs due to their focus on thinness and compact internal space.

From the perspective of downstream applications, the Internet, government, services, telecommunications, finance, manufacturing, education and other fields are areas with greater demand for computing power. Among them, the Internet field accounts for 49.7% of the overall computing power; followed by government, services, telecommunications, and finance, accounting for 12.0%, 10.0%, 7.2%, and 5.8% respectively.

From the perspective of GPU application terminals, my country's two main terminal application scenario markets of mobile GPU and desktop GPU continue to grow. According to relevant market statistics, the global GPU market has reached US$44.8 billion. In 2022, my country's GPU market size will reach US$8.36 billion, and in 2023, my country's market size will be approximately US$11.1 billion. With the development needs of AI, the general-purpose GPU market is growing rapidly.

With the expansion of the multi-field application-driven market, GPU companies have also developed rapidly. Among them, Xiangdi has passed mainstream production capacity and performance tests, and its product performance has reached the international first-tier level. Although there are many GPU companies in China, there is still no clear market structure.

1. Graphic rendering IP authorization and core IP self-research capabilities

GPU IP refers to the relevant intellectual property rights of the GPU. It includes various forms of intellectual property rights related to graphics processors, such as invention patents, trademarks, industrial designs, and geographical indications. GPU IP is mainly used to handle computationally intensive and highly parallel computing tasks, such as image and graphics processing, and can also refer to the design or architecture of the entire graphics processor hardware.

IMG is currently a leading company in the field of GPU IP. IMG has the world's leading complete mobile GPU IP, and its graphics and image parts are at the world's leading level, and it is a technological blank field in China.

However, IMG IP is mobile-oriented and has limitations in computing power: As a low-power mobile IP, IMG IP's main value is to empower desktop GPU graphics rendering functions, and its shader computing power is relatively weak. Although there are many GPU companies developing self-developed IP in the current market, self-developed IP may have technical risks, long research and development time, and high investment costs. IMG IP is the most mature and well-known in the field of mobile graphics rendering. So there is less chance of problems occurring.

Therefore, companies that can obtain permanent IP authorization from IMG in the future and have the ability to digest and absorb IMG IP and further innovate to form complete desktop GPU IP and other core IP self-research capabilities will have a competitive advantage.

2. Advanced microarchitecture is the core of GPU performance

Micro Architecture, also known as microprocessor architecture, is a hardware circuit structure used to implement instruction execution. A specific instruction set can be executed in different microarchitectures, but the technical effects during execution are different due to different design purposes. The GPU microarchitecture includes stream processors, rendering cores, double-precision floating-point computing units, special computing units, streaming multiprocessors, texture processors, graphics processors, and stream processor arrays.

With the expansion of downstream application fields, being able to adapt and be compatible with the market's mainstream operating systems, CPUs, and AI frameworks to meet the needs of various customers will become the mainstream trend in the future development of GPU enterprise intelligent ecosystems.

Global GPU chips are monopolized overseas, and overseas countries implement technical blockades on my country's high-end chips.

From the perspective of the global GPU chip industry chain, industry giants are mostly concentrated in the United States, Europe, Japan and South Korea. They are at the core of each link in the industry chain and play a decisive role in the global GPU industry.

In the independent GPU market, it is mainly monopolized by three companies: NVIDIA, AMD and Intel. Their global independent GPU market shares in 2022 will be 88%, 8% and 4% respectively. From the perspective of the PC GPU market structure, NVIDIA, Intel and AMD will have a market share of 18%, 62% and 20% respectively in Q2 2022. Intel occupies the largest market share with its advantages in integrated graphics on the desktop.

In addition, the United States and other countries have been implementing strict export controls in order to safeguard their own interests and prevent countries they consider unfriendly from obtaining technological products. Its export control direction for China mainly involves emerging technology industries, national defense and military industries, universities and scientific research institutions. The "encirclement and suppression" of sensors, chips, measuring instruments and other related companies is a top priority.

The US's export restrictions on high-performance GPUs deviate from the principle of fair competition on the one hand, and violate international economic and trade rules on the other, hindering international scientific and technological exchanges and economic and trade cooperation. In the context of my country's rapid development in the AI ​​field, the demand for AI chips has further increased. Therefore, the strong demand for domestic replacement of GPUs also provides good development space for high-quality domestic GPU chips.

my country has formed a series of policies to promote the development of the GPU industry

In the context of the current intensifying trade war between China and the United States and the need to consider technological decoupling between China and the United States, the importance of the independent and controllable strategy deployed by our country has become further highlighted. Under the national policy of safe substitution in the information industry, the localization of important IT components has become an urgent need. Governments at all levels have also released huge domestic substitution market space, providing a broad development stage for localized products and manufacturers.

There is strong demand for domestic substitution of GPU products in my country

Through the country's strong support in the past few years and the efforts of related enterprises, my country has basically available alternatives in terms of core chips for information products such as CPU chips, storage chips (DRAM/Flash), and network chips.

But in terms of GPU chips, almost all of the GPU chips currently used in our country come from NVIDIA and AMD . Among them, mid-to-high-end products account for more than 70% of the market share of GPU chips, and only a few domestic companies can achieve mass production and benchmark. NVIDIA is a company producing mid-to-high-end GPU chips. The lack of domestically produced products that can be applied in batches has always been a gap in the field of GPU chips in my country.

my country attaches great importance to the development of the chip industry. As early as 2006, the State Council promulgated the "National Medium and Long-term Science and Technology Development Plan (2006-2020)". "Core electronic devices, high-end general chips and basic software products" ranked first among 16 major science and technology projects, and also Known as the "01 Special Project", domestic GPUs are among them. It is imperative for my country's IT industry to become independent and controllable.

The importance of localization of IT infrastructure GPU

Currently, most of the CPU, GPU and other chips and key services used in important domestic information systems and critical infrastructure rely on foreign countries. Among them, 50+% of the information system host equipment in the fields of government, finance, energy, telecommunications, transportation and other fields use foreign products. For example, more than 70% of the equipment in the basic network comes from the American company Cisco; from Microsoft black screen door, Iranian nuclear power plant, Intel chip vulnerability It can be seen from many network security incidents such as Apple's Siri listening door and so on that Internet security is not only related to the safety of citizens' personal property, but also related to national economic development and national defense security.

As my country's IT infrastructure has been monopolized by overseas giants for a long time, there are certain risks in data security, network security and other aspects, especially involving party and government offices, state-owned enterprises, data centers and other fields, which are crucial to network security and data protection. . Therefore, it is necessary and urgent to take the lead in localizing IT technology software and hardware in core areas such as the information and innovation industry.

On January 29, 2024, 7 departments including the Ministry of Industry and Information Technology, the Ministry of Education, the Ministry of Science and Technology, the Ministry of Transport, the Ministry of Culture and Tourism, the State-owned Assets Supervision and Administration Commission of the State Council, and the Chinese Academy of Sciences jointly issued the "Implementation Opinions on Promoting Future Industrial Innovation and Development" , proposed that my country needs to accelerate breakthroughs in technologies such as GPU chips, clustered low-latency interconnection networks, and heterogeneous resource management, and build ultra-large-scale intelligent computing centers to meet the needs of large model iterative training and application reasoning. Driven by both policies and the market, domestic GPU manufacturers' product iteration speed and research and development speed will be further improved.

In the era of artificial intelligence, whether it is AIGC, GPT or the digital economy, the core technical facility behind it is the computing network. Computing power has become a key productivity factor in the digital economy era. China and the United States lead the world in terms of computing power. In 2022, the total computing power of global computing equipment will reach 906EFlops, with a growth rate of 47%. In 2022, global computing power competition will intensify. The shares of the United States, China, Europe, and Japan in the global computing power scale will be 34%, 33%, 17%, and 4% respectively. my country’s computing power scale ranks second in the world, led by GPU. Mainly realizes data center computing acceleration, with a market share of nearly 90%.

In the future, as the demand for intelligent computing power increases, my country's data centers will have strong demand for server GPUs. At present, the "Eastern Data and Western Computing" project has entered the comprehensive construction stage from system layout. It is expected that by the end of the "14th Five-Year Plan" period in 2025, the scale of data center racks will be increased to 14 million, and the total scale will be two in 2022. times, with a total incremental investment of approximately 700 billion yuan. It is expected that by 2028, the number of data center racks in my country will reach 23.39 million, with a five-year compound growth rate of 19.42%.

Assuming that the number of servers is calculated based on a five-year compound growth rate of 20%, the number of new servers in 2024, 2025, 2027, and 2028 will reach 3.8 million, 4.56 million, 5.472 million, 6.566 million, and 7.88 million. Each server is equipped with 4-8 GPUs (calculated as 6), and the price of a single GPU is 5,800 yuan. In the future, the annual incremental GPU market size of data center servers will be close to 200 billion yuan.

Currently, judging from the amount of data generated and processed by devices online on the Internet of Things, the speed and computing power of data centers are still far from enough, so edge computing was born. my country's edge computing industry is in a period of rapid development. my country's edge computing market will reach 70 billion yuan in 2022 and 91 billion yuan in 2023, a year-on-year increase of 30%. It is expected that the size of my country's edge computing market will reach 259.91 billion yuan by 2027.

The growth of the edge computing market has also driven the demand for servers. In 2022, the size of my country's edge server market will reach 45.5 billion yuan. According to the development needs of the edge computing industry, it is expected that the size of my country's edge computing server market will reach 1,468.464 billion yuan by 2027.

Generally speaking, an edge server needs to be equipped with 8 GPU cards. The price of a single chip is about 5,800 yuan to more than 10,000 yuan, accounting for 40% to 60% of the cost of the edge server. According to the size of the edge server market, it is predicted that the size of my country's edge server GPU market will reach 734.2 billion yuan in 2027.

AIGC's demand for high-performance computing has made AI servers the fastest-growing business segment in the server field. my country is also one of the major markets for AI chips. It is conservatively estimated that the size of my country's AI chip market will reach US$43.1 billion in 2028, with a five-year compound growth rate of 20%.

Among them, the demand for inference servers is increasing with the implementation of large AI models. It is expected that the market size of my country's AI inference servers will reach 95.6 billion yuan by 2027, of which 90% are GPU inference servers. According to the fact that a single inference server uses 8 GPU accelerator cards, accounting for 60% of the overall server cost, my country's inference server GPU market size will reach 64.5 billion yuan by 2027.

As ChatGPT triggers a craze for large AI models, NVIDIA, as a leader in accelerated computing, is currently in short supply of high-performance chips, and it generally takes at least 6 months to get products; coupled with US chip export restrictions. Facing the huge market demand for my country's AI chips, the domestic GPU market is promising.

After the epidemic, the orderly recovery of domestic social production and life and the accelerated pace of consumption recovery have driven the rebound and growth of the e-sports market. The compound annual growth rate of China's overall e-sports market size from 2023 to 2025 is expected to be 7.2%. In 2022, China's overall e-sports market size will reach 157.9 billion, and the number of e-sports users will reach 504 million yuan, further promoting the market development of desktop GPUs.

All calculations in cloud games are performed on cloud servers, and input instructions from players are accepted through the Internet. At the same time, the final screen results after processing are displayed on the player's front-end device. Therefore, a high-performance GPU needs to be installed on the server for cloud rendering, which will place higher requirements on GPU performance. China's cloud gaming market will be worth about 3.4 billion yuan in 2021, a year-on-year growth of 54.5%. It is expected that the overall compound growth rate of the cloud gaming market will reach 76.6% from 2022 to 2025.

Realizing the metaverse is inseparable from the technology of "digitizing the physical world." Presenting the physical world digitally puts higher demands on the GPU. Real-time rendering, physical rendering, and photorealistic rendering will all be completed by higher-performance GPUs. In the future, GPU graphics processing units will play an important role in how data is managed in the Metaverse.

In 2015, China issued the "Thirteenth Five-Year Plan for the Development of the Automotive Industry," setting goals for the development of intelligent connected cars. By 2020, my country will initially establish a standard system for connected cars that can support driving assistance and low-level autonomous driving. Automakers began integrating L2 autonomous driving in 2015. About 8% of cars sold globally in 2020 will be equipped with L2 and above autonomous driving functions. By 2025, about 30% of cars sold globally are expected to support L2 or above autonomous driving functions.

The global GPU industry market size will reach US$59.5 billion in 2023, and my country's GPU market size will be US$11.1 billion, making it the core market for GPUs. New technologies such as artificial intelligence, data centers, blockchain, and the Metaverse are emerging one after another, driving the GPU market to continue to develop rapidly.

However, the global GPU field is currently mainly occupied by manufacturers such as NVIDIA and AMD. In the face of environments such as the Russia-Ukraine war, information security, and the Sino-US trade war (technical blockade), key areas such as the party and government, state-owned enterprises, and data centers need to further ensure the independent controllability of national information security and core technologies.

The domestic replacement of GPUs in my country's Xinchuang, Xinchuang extension and other fields is at an early stage and is highly certain, which brings development space for domestic GPU companies. It also provides opportunities for domestic GPU companies to seize market share, further enhance product influence, and build a computing ecosystem. .

In addition, with the rapid development of many downstream fields such as artificial intelligence and the Internet of Things, the demand for GPUs in my country's data centers, cloud manufacturers and intelligent computing fields is also extremely strong. In the context of the Sino-US trade war and unstable global chip supply, my country has developed the AI ​​industry and formed an independent and controllable chip supply system, providing domestic GPUs with opportunities to enter the market.

Overall, the domestic GPU market has a vast space, and short-term market demand is more inclined to a "usable" state. When products become more mature in the future, companies that master core IP and architecture, have high product cost-effectiveness and can stably supply in batches will be more competitive. force.