A few days ago, the AI artificial intelligence special promotion meeting held by the State-owned Assets Supervision and Administration Commission (SASAC) became popular all over the network and received widespread attention from the industry.
The meeting required that enterprises should take the initiative to embrace the profound changes brought about by artificial intelligence and put the acceleration of the development of a new generation of artificial intelligence in a more prominent position. The meeting emphasized that "it is necessary to consolidate the basic foundation for development, concentrate the main resources on the most needed and advantageous areas, and accelerate the construction of a number of intelligent computing centers"; "Strengthen demand traction, accelerate the empowerment of key industries, build a number of industrial multimodal high-quality datasets, and create a large model empowerment industrial ecology from infrastructure, algorithm tools, intelligent platforms to solutions."
At the meeting, 10 ** enterprises signed a proposal, saying that they would take the initiative to open artificial intelligence application scenarios to the society. As an important pillar of the national economy, the layout and development of central enterprises in the field of artificial intelligence have attracted much attention. The accelerated deployment of artificial intelligence by central enterprises will bring new development opportunities to related industries, and also reflect the country's great attention to artificial intelligence and the general trend.
With the rapid development of global AI technology, intelligent computing power has become a key element to promote economic growth and social progress. As one of the world's largest economies, China is actively responding to this technological change by strengthening the research and development and application of intelligent computing power to enhance the country's competitiveness.
Before gaining an in-depth understanding of China's layout in the intelligent computing market, let's first understand what is "intelligent computing power"? When do we come into contact with intelligent computing power in our daily lives?
What is intelligent computing power?
According to the definition of "China Computing Power *** 2022", computing power is mainly divided into four parts: general computing power, intelligent computing power, supercomputing computing power, and edge computing power. The general computing power is mainly based on the computing power output by the CPU chip; Intelligent computing power is mainly based on the artificial intelligence computing power output by GPUs, FPGAs, AI chips, etc.; Supercomputing power is mainly based on the computing power output by supercomputers; Edge computing power is mainly based on providing users with real-time computing capabilities nearby, which is a combination of the first three.
From a conceptual point of view, it may seem a little far away, but in fact, intelligent computing power has penetrated into all aspects of our lives. If we talk about special effects, rendering and face recognition in Spring Festival movies, or applications such as intelligent customer service and voice translation that we use daily, these are inseparable from the support of intelligent computing power.
If artificial intelligence is an accelerator for innovation, then the intelligent computing center can provide support for all kinds of technological innovation.
How much intelligent computing power does China need due to the computing power gap caused by large models?
In 2023, many large-scale model products will be released, and the AIGC technology based on the large model has shown far exceeding expectations in text generation, knowledge answering, image generation, logical reasoning, etc., attracting a large number of users and market attention.
According to incomplete statistics, as of October 2023, there are 254 large-scale model innovation entities in China, distributed in more than 20 provinces, cities and regions, of which 122 are in Beijing, ranking first in the country. As of January 2024, 38 large models have been filed and launched, accounting for nearly half of the country. Released Wenxin Yiyan 40, the scale of users exceeds 100 million, and the average daily call volume reaches 10 million; Zhipu AI has developed the most advanced base model GLM4, and the open source version has exceeded 10 million times in the world, which is currently the domestic large model with the highest open source influence; Baichuan Intelligent released the large model baichuan2 that realizes the world's longest context window, and the open source version has exceeded 6 million times in the world in just four months; The Institute of Automation of the Chinese Academy of Sciences released the world's first 100 billion parameter multi-modal large model Zidong Taichu 20。In terms of deep learning frameworks, PaddlePaddle's popularity and usage rate are in the first echelon in China, with a domestic market share of nearly 36%, and as of the end of December 2023, 10.7 million developers have been gathered, serving 2350,000 enterprises and institutions.
In practical applications, the "digital human" effect created by the production method of automatically generating content using AI technology is comparable to that of real people. Artificial intelligence** protein structure brings new research methods to basic research; AI-powered chatbots are able to learn and understand human language and engage in conversations with humans; The Pangu Meteorological Model developed by HUAWEI CLOUD is ...... in predicting typhoon trajectories and landfall times
Behind the vigorous development of the visible AIGC is the support of invisible intelligent computing power. The explosion of large models has set off a new round of AI boom, and has also changed the demand and pattern of intelligent computing power.
The Action Plan for the High-Quality Development of Computing Infrastructure, jointly issued by the Ministry of Industry and Information Technology, the Cyberspace Administration of China, the Ministry of Education, the National Health Commission, the People's Bank of China, and the State-owned Assets Supervision and Administration Commission, proposes that by 2025, China's computing power will exceed 300eflops, and intelligent computing power will account for 35%.
At this point, some people may ask: what is FLOPS? What level is 300eflops?
FLOPS is a unit of computing power, which measures the number of floating-point operations per second of computing resources, and is the English abbreviation of floating-point operations per second. It is often used to estimate the performance of a computer, especially in the field of scientific computing, where a large number of floating-point operations are used, such as training and inference for machine vision-related image processing.
flops are preceded by a letter to indicate a larger unit of hash power:
One mflops (megaflops) is equal to one million (= 10 6) floating-point operations per second.
One GFLOPS (GIGAFLOPS) is equal to one billion (= 10 9) floating-point operations per second.
One tflops (teraflops) is equal to one trillion (= 10 12) floating-point operations per second.
One pflops (petaflops) is equal to one quadrillion (=10 15) floating-point operations per second.
One eflops (exaflops) is equal to 10 billion (= 10 18) floating-point operations per second.
Here are a few examples for easy understanding: when training an AI model with a large number of data samples, the computing power required for one training time varies from 2 to 19 pflos depending on the data scale, detection effect, and model type. In the inference process of face and language recognition, the demand for computing power may range from 10 GFLOPS to 64 TFLOPS depending on the recognition accuracy and the number of concurrency, and the computing power requirement for intelligent driving to complete functions such as environment perception, decision-making and obstacle avoidance, and self-vehicle positioning is about 8 TFLOPS.
As of the end of June 2023, the total scale of data center racks in use in China exceeded 7.6 million standard racks, and the total scale of computing power reached 197eflops, ranking second in the world.
Driven by the demand for large models, intelligent computing center projects have sprung up.
More than 30 cities have rushed to build intelligent computing centers
With the concentrated explosion of downstream computing power demand and the promotion of "Eastern Data and Western Computing", ******, operators, and Internet enterprises at all levels have started the construction plan of intelligent computing centers. According to the "Guidelines for the Innovation and Development of Intelligent Computing Centers" jointly issued by the State Information Center and relevant departments, there are currently more than 30 cities in the country that are building or proposing to build intelligent computing centers, including the Beijing-Tianjin-Hebei Big Data Intelligent Computing Center and the Changsha 5A-level Intelligent Computing Center.
The main enterprises in the construction of the intelligent computing center include the three major telecom operators and some Internet enterprises. The intelligent computing center promoted by operators has certain public service attributes and has become a good supplement to the construction of computing infrastructure led by the world. Internet companies represented by Alibaba, and Tencent have also built intelligent computing centers to promote their own business development and better promote the implementation of customer artificial intelligence scenarios.
As of August 2023, China's intelligent computing centers that have been put into operation or under construction are distributed in Beijing, Shanghai, Nanjing, Hangzhou and other cities. From the perspective of regional distribution, China's intelligent computing centers are concentrated in the eastern and central regions. Among them, the number of intelligent computing centers in the eastern region reached 25, accounting for 625%, ranking first, mainly in the Beijing-Tianjin-Hebei and Yangtze River Delta regions; The central region accounted for 175%, ranking second; The number of intelligent computing centers in the western and northeastern regions accounted for 125% and 75%。
It is worth noting that Beijing is one of the main regions that is the center of intelligent computing. Beijing has laid out intelligent computing centers in Haidian, Chaoyang, Jingkai Zone, Shijingshan, Mentougou, Daxing, Fengtai and other districts, and has built a total scale of intelligent computing power of about 5,000p.
Recently, Shijingshan District Beizhong Science and Technology Cultural Industrial Park of the four-span plant has started the hot renovation construction, after the completion of the initial will have 610p computing power, equivalent to the computing power of 300,000 high-performance computers, for an artificial intelligence model in 30 seconds to complete the learning and recognition of nearly 10 million **, is expected to be completed in October this year.
At the beginning of the year, there were frequent actions
At the beginning of the new year, major operators have taken active action to strengthen their layout in the field of intelligent computing centers.
On January 8, China Mobile Intelligent Computing Center (Wuhan) opened for operation in Wuhan Future Science and Technology City, with a service capacity of 1,500 pflops, and plans to expand to 6,800 pflops by the end of this year, becoming the largest intelligent computing center in central China.
On January 22, Shanghai Telecom lit up the "Large-scale Computing Cluster and Artificial Intelligence Public Computing Service Platform" in Shanghai, planning to build 15,000 cards in Shanghai in the first half of 2024, with a total computing power of over 4,500p, of which a single pool will build a new domestic computing power of 10,000 cards, which is expected to become the first ultra-large-scale domestic computing power liquid-cooled cluster in China.
On January 30, the inauguration ceremony of China Unicom's Artificial Intelligence Innovation Center was held in Beijing. It is worth noting that on November 24, 2023, China Unicom's Yangtze River Delta (Wuhu) Intelligent Computing Center project officially started.
It can be seen that the three major telecom operators of China Telecom, China Mobile, and China Unicom are focusing on the construction of "Eastern Data and Western Computing" data center clusters, making every effort to promote their respective related projects, accelerating the construction of national computing power center cities and intelligent computing centers, and promoting the deep integration of the digital economy and the real economy.
AI servers are the most critical equipment in the construction of intelligent computing centers
AI server is the largest and most critical equipment in the construction of intelligent computing center.
According to the latest report from market research institute IDC, from the first half of 2022 to the first half of 2023, China's AI server market size has grown by 54%, of which GPU servers are still dominant, accounting for 92% of the market share, reaching $3 billion. At the same time, non-GPU-accelerated servers such as NPUs, ASICs, and FPGAs accounted for 8% of the market share at a year-on-year growth rate of 17%, reaching $200 million.
In the first half of 2023, from the perspective of manufacturer sales, Inspur and Xinhua.
3. Ning Chang ranks among the top three, occupying more than 70% of the market share; From the perspective of server shipments, Inspur, Kunqian, and Ningchang rank among the top three, accounting for nearly 60% of the market share.
AI servers rely on high-performance chips. The computing power gap faced by the Chinese market has brought new opportunities for domestic chip development. China's local AI chip manufacturers are in a stage of rapid growth and have made remarkable achievements, attracting a lot of investment and attention. These companies have certain strength and competitive advantages in AI chip design, algorithm optimization, and manufacturing. In addition, China's policy support has also played an important role in promoting it. In the first half of 2023, the market size of China's acceleration chips will exceed 500,000. From a technical point of view, GPU cards account for 90% of the market share; From a brand perspective, China's local AI chip brands shipped more than 50,000 pieces, accounting for about 10% of the entire market.
From the perspective of the leading providers of AI acceleration chips required for domestic AI servers, domestic cloud service providers such as Alibaba (Hanguang series), (Kunlun series), and Huawei (Ascend series) have self-developed cloud AI acceleration chips. There are also Cambrian (Siyuan series), Haiguang Information (deep calculation series), Suiyuan Technology, Tiantian Zhixin, Biqian Technology, Moore Threads, Muxi, etc. In addition, Jingjiawei and Loongson Zhongke are also developing cloud AI acceleration chips.
Judging from the data in the first half of 2023, the localization rate of AI server chips in China has decreased compared with last year, and the proportion has dropped from about 15% last year to about 10%. This is mainly due to the significant increase in demand for high-end training servers, while the performance of domestic chips is difficult to keep up.
In order to further improve the performance of domestic AI servers, it is not only necessary for chip manufacturers to continue to innovate in technology to improve the performance and stability of chips, but also for these chip manufacturers to have an in-depth understanding of market demand and develop chip products that are more in line with actual application scenarios. At the same time, first-class enterprises and research institutions should also increase investment to support the development of the domestic chip industry and provide it with more R&D resources and market opportunities.
How to make the intelligent computing center really use?
After the completion of the intelligent computing center, how to make it play a greater role in the operation process is still a very critical issue.
At present, the intelligent computing center has gradually empowered the regional industrial exhibition, but it is worth noting that it still faces many challenges in the integration of multiple computing power, upstream and downstream collaboration, construction and application linkage, energy consumption, and use.
For example, general computing power and dedicated computing power have yet to be integrated. In different scenarios, such as autonomous driving, smart healthcare, and smart cities, the computing power requirements are different. A single computing power solution is difficult to meet the needs of multiple computing power, and cannot take into account multiple industries and fields.
Insufficient computing power, algorithms, and data collaboration. In the intelligent computing center built over the years, different chip platforms, algorithm models, databases, and application levels are in a state of vertical integration, and the compatibility of software and hardware needs to be improved.
Investment, construction and operation need to be linked. The investment, construction and operation of intelligent computing centers are often the responsibility of different entities. In the early stage, the construction unit often has insufficient investment in the mode and service standard of post-construction operation, and there is a phenomenon of separation between the head and the tail, and the construction and operation are separated, which affects the customer experience.
High carbon emissions and energy consumption. For example, OpenAI's ultra-large-scale pre-trained model GPT-3 consumes 190,000 kWh of electricity to train, which is equivalent to 228 times the per capita electricity consumption in 2021.
Investment costs and applications are to be regulated. The investment cost of some intelligent computing centers is as high as 500 million to 600 million yuan per 100p semi-accurate computing power, which is far higher than the normal market**, and the cost of use is also high, such as the training cost of GPT-3 large model is conservatively estimated to be more than 12 million US dollars.
The construction of intelligent computing centers needs to combine the construction foundation and local or regional industrial characteristics, classify and guide policies, reconstruct and parallelize, and develop intelligent computing centers that are compatible with the digital economy. It is also necessary to choose a reasonable construction and operation mode to achieve long-term operation, promote orderly layout, and ensure the maximization of the economic and social benefits released by the intelligent computing center.
Nowadays, the development of China's intelligent computing center industry is overcoming 1The challenge of the 0 era enters 2In the 0 era, China has always adhered to the principle of integrating computing power, synergy between software and hardware, integration of construction and operation, low-carbon energy consumption, cost optimization, demand traction, security and trustworthiness in the construction of intelligent computing centers, and steadily promoted the development of intelligent computing centers.