Domestic GPU chips urgently need to break through, and large models are hot promoting the demand for computing power. Artificial intelligence | Report | Domestic
Since the explosion of ChatGPT, there has been a craze for AI big models in China. Due to its massive training tasks requiring a large amount of computing power, GPU based computing power supply has become a key infrastructure for the development of the large model industry. However, a recently released report pointed out that the key core technologies supporting generative artificial intelligence in China's computing infrastructure are still under human control, and domestic GPU chips urgently need to be broken through.
The Beijing Blockchain Technology Application Association and Social Science Literature Publishing House jointly released the "Metaverse Blue Book: China Metaverse Development Report" on the 6th. A report titled "Research on the Development and Industrial Application of AIGC Technology" points out that the development of generative artificial intelligence is driving up the demand for computing power. GPU chips are currently the main hardware of generative artificial intelligence, with a global GPU chip market size of 33.47 billion US dollars in 2021. By 2025, the global GPU chip market is expected to exceed $400 billion.
The report's authors, Zhang Han and Wu Zhouming from the National Industrial Information Security Development Research Center, pointed out that in the global GPU market, two American companies, NVIDIA and AMD, have a market share of up to 96%. Data shows that in the third quarter of 2022, NVIDIA held 88% of the market share in the single chip GPU market, while AMD held 8% of the market share.
The report suggests that due to Nvidia and AMD having a complete product line and ecosystem in AIGC, their leading position in AI computing capabilities will continue for some time. In terms of hardware efficiency, training a ChatGPT requires support from 10 domestic data centers with a computing power of 500P, while Nvidia's latest GPU can increase the model training speed by 30 times.
The research companies for domestically produced GPUs mainly include Haiguang Information and Cambrian. From the technical parameters disclosed by Haiguang Information, there is still a certain gap between domestic high-end GPUs and foreign large companies in terms of graphics storage frequency and bandwidth. But in some specific application environments, the "Deep Computing One" of Haiguang Information can already be comparable to similar products in the world.
In the current complex international relations and limited chip imports, the development potential of the domestic artificial intelligence chip market is enormous. From 2020 to 2023, the compound annual growth rate of domestic AI chips reached 95.86%, and it is expected that their market size will exceed 130 billion yuan.
The report also points out that in terms of talent supply, AIGC technology with algorithm innovation as its core requires a large number of high-end R&D talents, and the talent resources related to new generation information technology such as AI in the United States are several times that of China. Our country has a large reserve of undergraduate talents in artificial intelligence, but there is still some loss and insufficient reserves.
The report suggests that domestic talent cultivation urgently needs to keep up with the forefront development direction of artificial intelligence such as ChatGPT, establish a talent support system that matches the artificial intelligence industry, and encourage universities and institutions to vigorously cultivate composite talents in the field of AIGC. It is urgent to optimize the system for introducing foreign-related talents, improve the accuracy of attracting high-end artificial intelligence talents, and strengthen the "source of vitality" for the innovative development of artificial intelligence in China, such as AIGC. In addition, it is recommended to build and improve the innovation ecosystem of the artificial intelligence industry, accelerate the transformation of AIGC achievements and industrial promotion.