Xinhua Finance and Economics, Shanghai, July 25 (Reporter Gao Shaohua) With the rapid development of artificial intelligence and multimodal large models, the demand for computing power has surged, and the world has ushered in a boom in the construction of intelligent computing centers. Experts said that at present, the construction of domestic computing infrastructure is in full swing, and domestic AI chips are facing broad application prospects, which will provide a solid computing power foundation for the digital economy.
The construction of intelligent computing centers has accelerated, and domestic AI chips have ushered in development opportunities
Since the release of ChatGPT in 2022, general-purpose large models have sprung up, putting forward very high requirements for AI infrastructure, which is manifested in the demand for computing power. In the next few years, the development of the entire AI industry will focus more and more on the GPU computing base. Computing power will become the infrastructure and a vital enabler for digital transformation and upgrading.
In the past two years, the global demand for computing power has continued to grow rapidly. According to TrendForce, AI server shipments in 2023 will be nearly 1.2 million units, up 38.4% year-on-year, accounting for nearly 9% of total server shipments, and is expected to account for 15% by 2026. From 2022 to 2026, the compound annual growth rate of AI server shipments will be about 22%. The International Data Corporation (IDC) predicts that the global AI chip market is expected to reach $72.6 billion by 2025. In 2023, mainland AI chip shipments will increase by 22.5% year-on-year.
The mainland's computing infrastructure is also accelerating its implementation. In 2023, the "Efficient and Quality Development Plan for Computing Infrastructure" issued by the Ministry of Industry and Information Technology and other six departments proposes that the scale of computing power in mainland China will exceed 300 EFLOPS in 2025, and the proportion of intelligent computing power will reach 35%. According to data recently released by the National Bureau of Statistics, as of the end of May, there were more than 10 intelligent computing centers with high-performance computer clusters planned across the country, and intelligent computing power accounted for more than 30% of the total scale of computing power.
Tang Wenkan, deputy director of the Shanghai Municipal Commission of Economy and Information Technology, said at the 2024 World Artificial Intelligence Conference that the innovation system of artificial intelligence large models is accelerating its evolution and gradually entering a new stage of comprehensive application in the industry, and the importance and influence of artificial intelligence computing power base have been further highlighted, indicating that the golden age of general artificial intelligence may have arrived.
At present, Shanghai is seizing the opportunity of a new generation of artificial intelligence, and there are dozens of intelligent chips with multi-technology routes in mass production, and a number of domestic AI chip manufacturers have emerged that have formed a complete product line and solutions, and have experienced multiple iterations and commercial use of chips. In the face of the challenges of the era of large models, Shanghai will continue to consolidate the industrial foundation and promote the development of domestic AI chips.
Zheng Weimin, an academician of the Chinese Academy of Engineering and a professor at Tsinghua University, believes that the explosive demand for computing power of large models usually includes four processes: one is to write software, tune software, and optimize software in the model development stage, the second is to require a lot of computing power in the training stage, the third is to train in vertical fields in the model fine-tuning stage, and the fourth is to be able to process user requests in real time when it is really used in the inference stage. It is very important to develop artificial intelligence and build a domestic intelligent computing system, and building a good software ecosystem can reduce the adaptation cost of large models to different AI chips.
Domestic AI chips are accelerating catch-up and building a computing power base in the AI era
Computing power is a key factor in promoting the transformation of AI technology and the development of the industry. At present, with the improvement of performance and ease of use, domestic computing power is gradually gaining the favor of domestic large models and artificial intelligence application enterprises.
Suiyuan Technology has started the construction and cooperation of intelligent computing centers in Chengdu, Sichuan, Yichang, Hubei and Qingyang, Gansu. Among them, Yichang Dianjun Intelligent Computing Center completed the construction and delivery of 300P domestic computing power in only one year, and realized full consumption online. "Computing power is of great significance to promote the rapid evolution of large models and the application of artificial intelligence, and Suiyuan Technology has completed many iterations of its products, and the scale of commercial landing is steadily advancing." Zhao Lidong, founder, chairman and CEO of Suiyuan Technology, said that he would seize the opportunity of the era of general artificial intelligence to jointly promote the development of local computing power.
The general computing application development and evaluation platform launched by Tiantian Zhixin has gathered more than 300 training model examples and more than 80 inference model examples to support various landing scenarios. Gai Lujiang, chairman and CEO of Tiantian Zhixin, said that in the context of a new wave of global AI caused by large models, high-quality computing power has become the core component of new quality productivity, and Tiantian Zhixin will deeply cultivate the field of high-performance general-purpose GPUs and work with partners to build a solid computing power foundation in the AI era.
Moore Threads' KUAE intelligent computing cluster solution has been scaled from the kilocalorie level to the scale of 10,000 calories. In the view of Moore Threads, the main battlefield of AI model training, Vanka is already the standard. As the amount of computing continues to rise, there is an urgent need for a gigafactory, a "large and general-purpose" accelerated computing platform, to shorten the training time and achieve rapid iteration of model capabilities.
Wu Chao, director of the China Securities Construction Investment Research Institute and chief analyst of the TMT industry, believes that in the field of artificial intelligence, the industry's concerns this year compared with last year's investment are more about computing power, and the capital expenditure on computing power is growing significantly in both North America and China. In addition to NVIDIA, the development of domestic computing power is also optimistic about the future.
Embrace the new era of intelligent computing and empower thousands of industries with intelligent computing power
While accelerating the improvement of the intelligence level and scenario generalization of algorithm models, the growing model parameters, massive corpus and extensive scenario requirements pose a huge challenge to the computing infrastructure of artificial intelligence, and how to make ultra-large-scale clusters have high efficiency and high cost performance at the same time has become a key problem that the industry needs to solve urgently.
Suiyuan Technology recently signed a strategic cooperation agreement with Qingcheng Jizhi, an artificial intelligence system software provider, to jointly develop high-performance system software solutions for large models with over trillion parameters and ultra-large-scale clusters. Zhao Lidong said that with the rapid development of large language models and the exponential growth of model parameters, the need to run complex models with the help of ultra-large-scale computing clusters is becoming more and more urgent.
After the past year's "100 model war" and "1000 model war", the number of large models in China is far more than United States, and how to make the large model truly commercialized is a problem that all large model companies urgently need to solve this year. Zhang Yalin, founder and chief operating officer of Suiyuan Technology, believes that 2024 will become the first year of large-scale model deployment, and how to create the ultimate commercial cost performance of computing power has also become an urgent problem for the industry to solve this year. This year, we can see that the business model of intelligent computing centers is also evolving rapidly, and computing power leasing and computing power consumption will drive the construction of China's intelligent computing centers.
In view of the development of the domestic computing industry, Liu Shanquan, CEO of Shanghai Intelligent Computing Technology Co., Ltd., believes that there is still a need for an exploration process with training clusters as the main business model. First of all, there must be a 10,000-calorie-level engineering and commercially available computing power cluster as the basic disk, and in this process, it is necessary to gradually solve the engineering problems in the construction of some ultra-large-scale computing power clusters. At present, from the perspective of the AI training computing power market, the overall demand tends to be stable, but there is still a demand for large-scale and ultra-large-scale single clusters. At the same time, with the gradual implementation of vertical industries, the demand for AI inference computing power may become the second curve of computing power growth.
As the core of the artificial intelligence industry, AI chips play a key underlying and fundamental role, and continue to empower thousands of industries. However, at present, the standardization of the AI chip industry in mainland China is still lagging behind the needs of technological development, and the development of the industry is facing deep-seated problems such as inconsistent technical standards and vicious competition in low-end homogenization.
The Ministry of Industry and Information Technology, the Cyberspace Administration of the People's Republic of China, the National Development and Reform Commission, and the National Standards Commission recently jointly issued the "Guidelines for the Construction of a Comprehensive Standardization System for the National Artificial Intelligence Industry (2024 Edition)", which points out that when improving the standards for smart chips, the general technical requirements related to smart chips should be standardized.
Chen Daji, vice president of China Electronics Standardization Institute, said that in the process of tackling key problems, developing and innovating breakthroughs in key industry application scenarios in the large model industry, it is necessary to take demand as the traction and standards as the guide, rationally deploy intelligent computing products, improve computing energy efficiency, accelerate scene innovation capabilities, jointly build a new ecology of intelligent computing, and jointly promote the high-quality green development of related domestic industries.
Editor: Hu Chenxi
Statement: Xinhua Finance is a national financial information platform undertaken by Xinhua News Agency. In any case, the information published on this platform does not constitute investment advice. If you have any questions, please contact customer service: 400-6123115
![Enterprise digital agent: Build an enterprise digital and intelligent AI foundation[fig]](data:image/gif;base64,R0lGODlhAQABAIAAAP///wAAACwAAAAAAQABAAACAkQBADs=)