AI Pushes Up the Total Computing Power Industry in China. How many steps is it from ＂computing＂ to ＂intelligent computing＂？

　　When Huang Renxun, the founder of NVIDIA, shouted "The tipping point of generative AI is coming" in May this year, a competition around the global computing industry chain was also going on.

　　"China’s computing industry has begun to take shape, and the output of computing products such as servers, computers and smart phones ranks first in the world. Judging from the total scale of computing power, it ranks second in the world. " On August 19th, Jin Zhuanglong, Party Secretary and Minister of the Ministry of Industry and Information Technology, said at the Computing Power (Infrastructure) Conference in China in 2023 that computing power has become the key productivity in the era of digital economy and an important cornerstone of the digital and intelligent transformation of the whole society, and it is necessary to speed up key technology research.

　　Enterprises in the industrial chain are also feeling the opportunities brought by this computing revolution. "Everyone is scrambling for the layout." A manager of Hyperfusion told reporters that the scale and speed of domestic (enterprise) investment and deployment are accelerating, whether it is intelligent computing or specific to large-scale model computing power.

　　AI computing power demand "jumps"

　　As an important productive force in the era of digital economy, the scale of China’s computing core industry has reached 1.8 trillion yuan in 2022.

　　According to the Evaluation Report of Global Computing Power Index in 2022-2023, every dollar spent on IT can boost the digital economy output of 15 dollars and GDP output of 29 dollars. In other words, the digital economy will grow by 3.6&permil for every 1 point increase in the country’s computing power index; , GDP will increase by 1.7‰ .

　　According to the latest data released by the Ministry of Industry and Information Technology on the 19th, up to now, the total rack size of data centers in use in China has exceeded 7.6 million standard racks, and the total computing power has reached 197 trillion floating-point operations (197EFLOPS), ranking second in the world. In addition, 130 trunk optical cables were built around the hub nodes of computing power, and the data transmission performance was greatly improved.

　　By the end of 2022, there were more than 6.5 million standard racks in use in China, with a total computing power of 180EFLOPS. In contrast, in the first eight months of this year, the two figures increased by 16% and 9.4% respectively. According to IDC data, influenced by AI, from 2022 to 2026, the compound annual growth rate of artificial intelligence computing power in China will reach 52.3%.

　　"The door to pattern reshaping has been opened, and the domestic computing industry is undergoing unprecedented major changes." Liu Hongyun, chairman and CEO of Superconfusion, used "jumping" to describe the current state of the industry. He believes that the big model is giving birth to more demand for AI computing power and entering the era of "intelligent computing".

　　"The parameters of large language models have grown from 100 million in 2018 to 100 billion in GPT-3 in 21 years, increasing by 1,000 times in five years. Correspondingly, the demand for computing power of these models has increased by 10 times every 18 months, which is 5 times that of Moore’s Law. In recent years, with the help of sparse computing MoE theory, a large language model with trillions or even trillions of parameters has emerged. " Liu Hongyun said at a partner summit that the rise of multimodal AI will bring more complex models and more huge computing power requirements. The large model is to AI as the earth is to all kinds of animals and plants, which greatly improves the speed and quality of AI development and application.

　　The industry where hyperconfusion is located is the "server" link in the computing power industry chain. At present, the company’s share in the industry has reached the top two, second only to Inspur.

　　For players in the computing power industry chain like hyper-fusion, trillions of parameter models are constantly emerging in the AI era, and the demand for diverse computing power is also growing.

　　According to the data provided by Tianyancha to reporters, by the first half of 2023, there were more than 20 financing events directly related to the "big model" and more than 40 patent applications related to the big model. In the era of big model represented by GPT, multi-modal AI technologies such as voice, picture and video have risen rapidly, shaping a wider data form.

　　As the AI model enters the industry, the computing power it brings will also be reflected in the fields of government affairs, industry, transportation, medical care and other industries. The reporter noted that since last year, Henan, Hangzhou, Chengdu, Wuhan, Shanghai, Ningxia and other places have successively introduced policies to support the development of computing power to promote the deep integration of technologies such as the Internet, big data and artificial intelligence with the real economy.

　　What are the key points? Where is the challenge?

　　However, while the computing power industry is developing rapidly, it is also facing risks and challenges, such as energy consumption and insufficient computing power.

　　According to statistics in the industry, it takes 14.8 days to train the GPT-3 model on 1000 NVIDIA V100 GPU. Under the condition that the PUE of the data center is 1.1, the total energy consumption will reach 1287MWh. Based on the per capita living electricity consumption level in China in 2021, the power consumption of a single large model training is equivalent to a person’s total living electricity consumption for four years.

　　In addition, there is still a gap between demand and supply in the computing market. According to the prediction of research institutions, the amount of newly generated data in the next three years will exceed the sum of the past 30 years. However, the total amount of data is increasing, and the proportion of data that is really effectively used is negligible. In key technologies, such as server chips, Intel (Intel), AMD and NVIDIA account for more than 85% of the domestic server chip market, and the supply of high-performance chips is insufficient.

　　"The change of computing power demand is also forcing us to go upstream, and the joint ecological partners will reshape the architecture around the server base." Zhang Xiaohua, President of Hyperfusion Global Marketing & Sales Service Department, told reporters that the most important thing in the computing industry is the consensus and promotion of eco-industrial chain partners.

　　"Ecology We have defined multiple dimensions, including sales, services, upstream suppliers, joint innovation lab, software service providers and industry standard organizations, and provided support for partner businesses in terms of systems, incentives, rights, support and services." Zhang Xiaohua told reporters that the "double-ecology" mode currently adopted internally. On the one hand, it cooperates with global suppliers of head parts and raw materials. On the other hand, the free combination of domestic hardware and software products is realized through its own operating system and virtualization technology.

　　In addition to deploying the software and hardware ecology, China manufacturers are also actively deploying the solution of computing power consumption, among which liquid cooling technology has become the direction of tackling key problems.

　　At present, Internet vendors including Ali and Tencent, server vendors such as Hyperfusion, Inspur Information and Dawning have successively invested in the construction of liquid cooling equipment. In order to solve the energy consumption problem, the three major operators plan to carry out large-scale application of liquid cooling by 2025, and more than 50% of data center projects will adopt liquid cooling technology.

　　"From the whole liquid cooling architecture to the realization of liquid cooling, and then to the most critical heat dissipation link involved in the liquid cooling transmission process, the technology has iterated to the fourth generation." Zhang Xiaohua told reporters that the rhythm of R&D is product generation, research generation and operation generation. At present, 10 XLab joint innovation lab have been established in conjunction with several industrial partners, covering key technologies at all levels, from materials to devices, from board-level components to equipment level, and from ecology to data centers.

　　"Technological breakthroughs are fundamental to the development of computing power, and it is necessary to closely track the global technological evolution and industrial development trends." Jin Zhuanglong said at the conference that it is necessary to strengthen systematic innovation and firmly grasp the leading role in development.

　　Previously, the Ministry of Industry and Information Technology planned to issue policy documents to promote the high-quality development of computing infrastructure, further strengthen the top-level design, enhance the ability of independent innovation, and enhance the comprehensive supply of computing power.

　　At this conference, Jin Zhuanglong emphasized that China’s computing power industry has begun to take shape, and high computing power chips have accelerated iterative upgrading, and a number of key enterprises in the industry have grown sturdily. (Next) We will carry out the "strong computing power" and give full play to the traction role of "chain owners" enterprises. Focusing on key links such as computing, network and storage, we will gather scientific and technological strength, increase investment in research and development, break through a number of landmark technical products and programs as soon as possible, and accelerate new ones.