The demand for computing power is exploding under the upsurge of AI large-scale models. How to play in Shanghai?

Source: The Paper

Author: He Liping

Original title: "Special Article｜The demand for computing power explodes under the boom of AI large-scale models: Lingang wants to build a tens of billions of industries, and SenseTime will be the "chain owner""

On January 24, 2022, SenseTime's artificial intelligence computing center AIDC, which was deployed in the Lingang New Area of Shanghai Free Trade Zone, officially launched operations. At that time, the artificial intelligence company may not be able to accurately predict that 2022 will become the first year of the so-called AIGC (AI Generated Content).

"Today, our Lingang AIDC has nearly 30,000 GPUs (graphics processing units), and our current computing power has reached 5,000 PetaFLOPS (1 PetaFLOPS is equal to 1 trillion floating-point operations per second). We believe that there will be better developer efficiency in the future, and they will be able to support more large-scale model computing power training with a scale of 100 billion." On June 2, "AI leads the era, computing power drives the future" - Lingang Xu Liru, chairman and CEO of SenseTime, said at the Intelligent Computing Conference in the new area.

According to Peng Mei News (the reporter learned from SenseTime, "there are still a lot of demands queuing up on the road." According to Yang Fan, the co-founder of SenseTime and the president of the large-scale device business group, artificial intelligence is very important for larger data, The pursuit of larger scale and greater computing power did not "begin today", "the entire history of artificial intelligence technology iteration and progress, we can regard it as a pursuit of 'violent aesthetics', and algorithms, computing power , The three elements of data are a technical iterative process in which quantitative changes produce qualitative changes."

The Lingang New Area, which focuses on cutting-edge industries, quickly responded to this new wave. On June 2, Wu Xiaohua, deputy secretary of the Party Working Committee of the Lingang New Area, released the "Action Plan for Accelerating the Construction of a Computing Industry Ecology in the Lingang New Area" at the above conference. Under the blueprint of the "Plan", by 2025, Lingang will become a computing power industry cluster with national influence, and the overall scale of the computing power industry including related hardware, software, applications, services, etc. will exceed 10 billion yuan.

Wu Xiaohua, deputy secretary of the Party Working Committee of the Lingang New Area, issued the "Action Plan for Accelerating the Construction of a Computing Power Industry Ecology in the Lingang New Area".

"We see that the era of AI explosion has come. AI has entered various fields of our production and life, so with the explosion of AI applications, it has actually driven the explosion of computing power demand." Regarding the above-mentioned "Proposal" ", Lu Yu, director of the high-tech department of the Lingang New Area Management Committee, told The Paper (including the media that Lingang already has a good advantage in the early stage, "that is, our computing power resources are very rich."

More importantly, when artificial intelligence companies choose whether to land in Lingang, computing power resources have become a particularly important decision-making factor.

Computing power is the energy of the new era, and success is not only about "violent aesthetics"

What is computing power? Xu Li believes that computing power is actually an expression of the capability of the entire model. "Computing power is equal to the parameters of the algorithm or large model, multiplied by the amount of data it processes. The larger the parameters in the era of large models, the greater the amount of data multiplied. , the greater the computing power required.” Computing power has become the energy source of the new era, “to some extent, computing power determines the competitiveness of the market.”

Yang Fan also mentioned that from last year to this year, a very popular concept in the field of artificial intelligence is called content generation. At the same time, everyone is familiar with a term that is a large model. In simple terms, this is a kind of "violent aesthetics". For example, the GPT-3 model uses more than 175 billion parameters and requires a high-performance processor to support training. Using V100 training consumes 10,000 cards and trains for 14.8 days , the overall computing power requirement is about 625 PetaFLOPS.

Yang Fan believes that this kind of "violent aesthetics" can also be understood as quantitative changes leading to qualitative changes. "In fact, from the first day of its birth to today, artificial intelligence has been pursuing greater intelligence through scale." He mentioned In fact, as far as the field of artificial intelligence is concerned, in the past 5-6 years, the consumption of computing power by the top artificial intelligence models in the industry has "doubled every 4-6 months, that is, It is said that it has doubled by nearly 300,000 times in the past few years."

Of course, "violence" and "intelligence" are not completely proportional. "Having greater resources and a larger scale is only a necessary but not sufficient condition." Yang Fan emphasized that the real support behind the "violent aesthetics" It is precisely the continuous optimization and improvement of each link that produces major technological innovations and achievements.

Take the data as an example, "The data used by GPT-4 is actually only 1% of all the data collected by OpenAI, because he found that when more data is fed to the robot, it may not be so smart. Effective and more valuable data is provided to this algorithm, and then a smarter brain can be created."

It believes that, at least today, the validity of data is far more important than the total amount of data. As for how to define effective data, "This actually requires a lot of hard work by data scientists. OpenAI actually let their best scientists do data, not algorithms as everyone thinks."

This optimization of each link also includes computing power. Why is no one using domestic chips for commercial large-scale training when Nvidia is out of stock? Why was Nvidia earning money for the first time after the latest wave came? The explanation behind these questions is, "It's not that we can generate the final value by piling up the computing power to a value. 1000 cards and 100 servers are put together to run the same task, which requires a lot of supporting software and communication. Network, it is a series of software and hardware joint optimization process, we have not done such work accumulation in the past, we need to make up lessons today.”

## Following the trend, Lingang speeds up the formation of a multi-computing power supply system

According to Wu Xiaohua, the computing power industry in Lingang New Area has been deployed in upstream software and hardware, midstream data centers, scheduling platforms, and downstream applications. At present, the total computing power of Lingang exceeds 3EFLOPS (FP32, 1 EFLOPS is equal to 10 billion per second 100 million floating-point operations), of which intelligent computing power accounts for nearly 80%, and the total computing power scale accounts for nearly 20% of Shanghai.

The above-mentioned "Proposal" proposes that by 2025, the new area will form a multi-computing power supply system based on intelligent computing power, basic computing power and super computing power. The total computing power exceeds 5EFLOPS (FP32), and the AI computing power The proportion reached 80%, the overall scale of the computing power industry (including related hardware, software, applications, services, etc.) exceeded 10 billion yuan, a public computing power service platform was built, the computing power trading mechanism was standardized, regional computing power scheduling was realized, and a nationwide Influential computing power industry clusters, build a batch of computing power demonstration application benchmarking scenarios.

"Intelligent computing power is what the most popular AI companies need at the moment. We also found that when AI companies land in Lingang, they no longer just focus on how much policy support and how much subsidies they will give them. They will pay attention to where they land. Here, can it solve his demand for computing power, because computing power is very scarce in the market now." Lu Yu mentioned this significant change.

According to SenseTime, as of May this year, SenseTime's large devices have served more than 40 core customers. "Especially under the wave of large-scale models, we now support more than 10 organizations to train their large-scale models in the intelligent computing center in Lingang." Yang Fan also mentioned.

Shenshi Technology, established in 2018, is one of the computing power demanders. The company's core team is led by E Weinan, an academician of the Chinese Academy of Sciences, and is the pioneer of the "AI+Science" scientific research paradigm. Its first "multi-scale modeling + machine learning + high-performance computing" paradigm has made a breakthrough. Unification of precision and efficiency in multiscale molecular simulations.

According to The Paper (reporter previously reported, Shenshen Technology has launched Lebesgue scientific computing platform, Hermite drug design platform and Bohrium micro-computing and design platform, etc. For example, in the field of medicine, Shenshen Technology has joined hands with many customers to combine physical modeling The computing paradigm of AI is more widely combined with preclinical drug research and development. Through Hermite Uni-FEP, Uni-Fold, RiD and other modules, the free energy perturbation theory, molecular dynamics, enhanced sampling algorithm and high-performance computing are combined to achieve accurate Predict protein structure and conformational changes, and efficiently evaluate the binding free energy of proteins and ligands with chemical precision, provide efficient and accurate theoretical guidance for drug developers, and improve the efficiency of drug design and optimization.

On December 29 last year, Beijing-based Shenshi Technology registered and established Shenshi Energy Biotechnology (Shanghai) Co., Ltd. in Lingang. Liu Huishi, vice president of government and enterprise affairs of Shenshi Technology, said in an interview with The Paper (reporter) that the company's deployment of a new generation of molecular simulation algorithm research and development center and AI-assisted drug design business center in Lingang is mainly because it sees that Lingang is developing vigorously. Computing power, "We have a demand for computing power in the process of training models. In addition, Lingang especially needs to vigorously develop localized computing power. We also want to contribute to this aspect."

From a business perspective, “We are mainly deploying our drug research and development business in Lingang, including the research and development of our own pipeline.” Liu Huishi mentioned that Shenshi Technology’s business is related to artificial intelligence and biomedicine in Lingang and even Shanghai. Such a leading industry has a positive and direct cooperative relationship. "We are willing to incorporate our research and development and products into the large ecological circle of Lingang."

The above-mentioned "Plan" also mentioned that the Lingang New Area has also formulated a series of safeguard measures, including strengthening talent protection, improving support policies, and promoting open cooperation. According to Lu Yu, if AI companies come to Lingang, they will give priority to providing intelligent computing power in Lingang, and at the same time, through the issuance of computing power vouchers and other forms, companies will be allowed to use computing power preferentially. "Even for key AI companies, the government will The cost of computing power can be directly subsidized by no more than 30%, and we will come out with these policies.”

It is worth noting that at the conference site, China Telecom Lingang public intelligent computing service platform and domestic GPU joint innovation base were also officially released. China Telecom established Lingang Computing Power (Shanghai) Technology Co., Ltd., which will carry out the construction of Lingang Computing Power Park, and will release 40,000 high-power racks suitable for intelligent computing and super computing in batches.

Tang Wenkan, deputy director of the Shanghai Economic and Information Commission, said that at present, the new generation of information infrastructure with "network as the foundation, data as the core, computing power as the key, and security as the bottom line" has become an important foundation for the construction of modern industries. Shanghai has proposed to build a modern industrial structure of "2+(3+6)+(4+5)", which puts forward higher demands on the construction of new information infrastructure represented by computing power.

And just on May 16, the Shanghai Municipal Economic and Information Commission announced the list of data center projects that passed the conformity assessment of the "Guidelines for the Construction of Data Centers in Shanghai", and supported a total of 16 projects, including projects located in Lingang. 2. "Up to now, our committee has supported 8 projects in the new area, including SenseTime AIDC, Youfu Network, and Information Flying Fish, with a total of 28,000 6kW standard cabinets, accounting for nearly 1/5 of the city's approved cabinets."

Tang Wenkan also put forward one of the suggestions, which is to use Lingang's abundant computing power resources to build public computing power services. "Currently, SenseTime AIDC in Lingang has connected to the public computing power service platform. I also hope that all units participating in the meeting today, especially telecom operators, will actively build extremely fast computing power in Lingang based on the characteristics of Lingang's network. Power-carrying network, helping to realize the ubiquitous network, ubiquitous computing power, ubiquitous intelligence, and promote computing power to become a public service like hydropower."

Established the Intelligent Computing Industry Alliance, SenseTime will be the leader of the industrial chain

Based on existing advantages and future needs, Lingang hopes to build a computing power industry alliance integrating upstream, midstream and downstream for coordinated and systematic development.

Lu Yu regards the supply of computing power in Lingang as the "middle section" of the entire industrial chain. One end provides computing power guarantee for AI companies landing in Lingang, and the other end involves "chips" that are extremely critical to computing power. , software, and system”, “we hope to have a demand side as well as such a platform side, then we will gather computing chip companies, software companies, and system companies here, and let them deeply participate in the construction process of such a system among."

Yang Fan also emphasized, "The development of all the achievements of the large model seen today is not only a miracle of violence, the improvement of technical value brought about by the continuous scale increase of the three elements of artificial intelligence, but also the basic research and development capabilities and systematic engineering. The in-depth combination of computerization capabilities, algorithm optimization, data sorting and selection, and platform computing power optimization and provision are often interconnected, and it is difficult to turn it into a separate link to do it alone."

He said that the important value of the intelligent computing power industry chain lies in, "Only when there are more companies on the chain, everyone can promote mutual exchanges and thinking, and carry out some cooperation in a deeper level, can we be able to develop such new key major technologies." In the tide, to achieve better technological progress and support."

At the conference site, the Smart Computing Industry Alliance in the new area was also formally established, with China Unicom as the first rotating chairman unit of the alliance. It is reported that in the future, China Unicom will establish the Yangtze River Delta Innovation Research Institute in the new area to further help the development of the intelligent computing industry in the new area.

The members of the Intelligent Computing Industry Alliance in the new area are represented by computing power providers such as intelligent computing power, basic computing power and super computing power centers, computing power chip companies such as GPU, FPGA, and ASIC, and computing power such as large models and AI for science. Demanding enterprises, a total of 25 enterprises, and China Institute of Information and Communications Technology East China Branch, Xi'an University of Electronic Science and Technology of China, and University of Electronic Science and Technology of China, a total of 3 universities and research institutes, will carry out resource sharing, technical exchanges and project cooperation in the future. SenseTime was awarded the "Chain Leader of the Intelligent Computing Industry Chain in the New Area".

GPU chip manufacturer Mu Xi said on the same day that three types of GPU products that meet the functions of AI inference computing, AI training/general computing, and high-performance rendering are applied in AI reasoning, AI training, data centers, metaverses, cloud games and other fields. It will empower the transformation and development of various fields.

Tang Wenkan also has high hopes for the establishment of the Intelligent Computing Industry Alliance in the Lingang New Area, "Relying on the main chain enterprises such as SenseTime, combined with their own advantages, to explore the synergy of all elements of the upstream and downstream of the industrial chain, and form a new explosive point of the digital economy. "

At the site of the conference that day, 12 companies jointly signed a collaborative procurement agreement for upstream and downstream enterprises in the smart computing industry in the new area. Lu Yu mentioned that the new area will also issue a positive list for collaborative procurement, "If companies purchase domestic GPUs and other upstream products in the process of building a localized computing power platform, we will give subsidies, which also encourages upstream and downstream companies to go more. Good cooperation."