With the development of Artificial General Intelligence (AGI), the demand for computing power required for automatic AI content generation technology is also rising.
According to the technical documentation of SORA, the first Wensheng ** model released by OpenAI, with the increase in the amount of training and computation, the quality of samples has improved significantly, which further proves that in the multimodal era, the upper limit of computing power demand is still unimaginable.
Guotai Junan believes that the release of the SORA model will further widen the computing power gap. The SORA large model continues the technical path of the previous GPT model "source data-transformer-diffusion-emergence", which means that its development and maturity also require massive data, large-scale parameters, and large computing power as the base, and because the amount of data required for training is much larger than the amount of data for text training, it is expected that the SORA large model will further widen the computing power gap.
Intelligent computing centers can drive the growth of AI core industries29 to 34 times.
Generally speaking, computing power is divided into three major sections: general computing power (basic computing power), intelligent computing power and supercomputing computing power. Among them, the basic computing power is mainly provided by CPU chip-based servers, which are used to support basic general computing such as cloud computing and edge computing. Intelligent computing power is provided by accelerated computing platforms based on AI chips such as GPUs, FPGAs, and ASICs, and is mainly used for AI training and inference computing. Supercomputing power is provided by high-performance computing clusters such as supercomputers, and is mainly used for computing in cutting-edge scientific fields.
OpenAI CEO Sam Altman said a few days ago that he is in talks with investors to increase global chip production capacity and expand its ability to support AI performance, and the project may need to raise up to $5 trillion to $7 trillion.
According to the "2022-2023 Global Computing Power Index Evaluation Report" jointly released by IDC and Inspur Information, the global AI computing market will increase from 195 in 2022$0 million to $346 by 2026$600 million, of which the generative AI computing market size will increase from 8$200 million to $109 by 2026$900 million. The share of generative AI computing in the overall AI computing market will increase from 42% grew to 317%。
What is the current situation of intelligent computing power construction in China?
According to the China Academy of Information and Communications Technology's "China Computing Power Development Index 2023", China's computing power will expand steadily in 2022, and intelligent computing power will maintain strong growth, with China shipping more than 20.91 million general-purpose servers and 820,000 AI servers in the past 6 years, and the total computing power of computing equipment will reach 302eflops, of which intelligent computing power will grow rapidly, with a growth rate of 72%, accounting for more than 59% of China's computing power. According to the Guidelines for the Innovation and Development of Intelligent Computing Centers (hereinafter referred to as the "Guidelines") issued by the State Information Center, during the 14th Five-Year Plan period, when the intelligent computing centers achieve 80% application level, urban investment in intelligent computing centers can drive the growth of the core AI industry29 to 34 times, driving the growth of related industries by 36 to 42 times.
An expert in the field of artificial intelligence told Yicai that there is still a gap between China and the United States in terms of intelligent computing capabilities, and the chip problem has also affected the intelligent computing center. Although China currently has advantages in the scale of computing power, there are still some gaps in how to make better use of computing power and how to improve the efficiency of computing power. "On the basis of the advantages of chips, the United States is also constantly expanding the scale of computing power, especially the scale of intelligent computing power. ”
What are the trends in the construction of intelligent computing centers?
Although the general computing power data center is currently the mainstream, the construction of intelligent computing centers is also being accelerated in various places.
In October 2023, the Ministry of Industry and Information Technology and other six departments issued the "Action Plan for the High-Quality Development of Computing Infrastructure", proposing that by 2025, the scale of computing power will exceed 300eflops, the proportion of intelligent computing power will reach 35%, and the balance and coordinated development of computing power in the east and west. The total storage volume exceeds 1,800 EB, the advanced storage capacity accounts for more than 30%, and the disaster recovery coverage rate of core data and important data in key industries reaches 100%.
Yu Xiaohui, president of the China Academy of Information and Communications Technology, pointed out in an article that with the development of the computing economy, the integration and innovation of computing technology and artificial intelligence have made the intelligent computing center a new infrastructure hotspot, that is, a center dedicated to artificial intelligence computing. As of March 2023, more than 30 cities in China are building or proposing to build intelligent computing centers.
Beijing, Shanghai, Guangzhou and other areas with developed digital economies have also introduced policies to promote the construction of intelligent computing centers. For example, Beijing proposes to "build a new batch of computing data centers and artificial intelligence computing centers"; Shanghai proposes to "propose the implementation of a large-scale model intelligent computing power acceleration plan; Build a municipal-level intelligent computing power overall scheduling platform, and build large-scale advanced computing power scheduling and supply capacity".
In terms of specific projects, on February 4, the lighting operation of the Shenzhen Open Intelligent Computing Center and the unveiling ceremony of the Shenzhen Smart City Computing Power Coordination and Scheduling Platform were held, which means that Shenzhen is accelerating the construction of a 100,000-calorie-level super computing power cluster. Based on the international mainstream intelligent computing chips, the Shenzhen Open Intelligent Computing Center can realize the training of hundreds of billions of large models with high complexity and high computing requirements.
The recently launched Beijing Artificial Intelligence Public Computing Platform (Shangzhuang), as an artificial intelligence infrastructure, will provide inclusive public computing services for all kinds of innovative entities. Among them, after the first phase of 500p (p is the quantitative unit of computing power, 1p is equal to 1000 trillion calculations per second) computing power is launched, it will provide inclusive computing services for universities, scientific research institutes, and small and medium-sized artificial intelligence enterprises in Beijing, alleviating the current situation of shortage of computing resources.
On January 22, at the "Computing Power Huimo Du 10 Gigabit Liansu City" artificial intelligence public computing service product launch conference, the "large-scale computing power cluster and artificial intelligence public computing service platform" was lit up, "This will be the largest operator-level intelligent computing center in the country." Gong Bo, Secretary of the Party Committee and General Manager of Shanghai Telecom, said that the scale of the computing power cluster lit up that day reached 15,000 cards.
In Wuhan, China Mobile's first batch of intelligent computing center nodes in China, China Mobile Intelligent Computing Center (Wuhan), opened for operation at the end of last year, using the all-optical base of the computing power network that has been put into production, with a "1-5-7-10" ultra-low latency circle covering central China and radiating the whole country. In 2024, the computing power scale of the intelligent computing center will reach 6,800 pflops, making it the largest intelligent computing center in central China.
Next, what are the trends and priorities of the construction of the intelligent computing center?
The above-mentioned "Guide" points out that at present, there are three independent investment models for the intelligent computing center: first, independent investment and construction, the construction funds come from local financial funds, special bond issuance, etc., and the ownership is owned by the first after completion; the second is the independent investment and construction of enterprises, which are invested in the form of enterprise alliances, a small number of enterprise alliances, and independent enterprises; The third is that universities or scientific research institutions independently invest in construction and provide free computing power support to teachers, students, and researchers, which has a lower construction cost and serves the university education scene.
According to the Guidelines, intelligent computing centers can adopt different construction models for different geographical conditions. Although Beijing, Shanghai, Guangzhou, and the economically developed and densely populated regions in the east have high demand for data storage and processing, they are faced with problems such as tight regional energy consumption indicators, high power costs, and limited development space, so the original data centers can be transformed and upgraded to meet the needs of national and local government services, major projects and key laboratories.
Shan Zhiguang, director of the Department of Informatization and Industrial Development of the State Information Center, also mentioned four aspects in a public speech when he elaborated on the development trend of the intelligent computing center, namely, the rapid expansion of the demand for intelligent computing power, the intelligent computing center is becoming the key to the development of general artificial intelligence, the gradual construction of a universal and inclusive service ecology, and the acceleration of the formation of a green and low-carbon development pattern. For example, the construction of perception, control, and interaction capabilities with multimodal fusion technology as the core is an important exploration direction for the realization of general artificial intelligence. The intelligent computing center must do a good job of carrying functions and face new artificial intelligence technology innovations such as large models.
Shan Zhiguang said in an interview that the current industry's understanding of the concept, operation mode and profit model of the intelligent computing center is not in place, resulting in the unscientific charging model of some projects and the little support for the application. It is hoped that the industry will pay attention to the rational planning and layout of the intelligent computing center, and carry out construction and development for the purpose of ultimate benefits and effects, such as whether it can drive industrial clustering and whether it can achieve economies of scale.