Chinese intelligent computing power service, will the plan be installed?

Benefiting from a new round of artificial intelligence boom, the stock price of US GPU chip manufacturer NVIDIA has tripled in less than 8 months; At the end of May, the company's market capitalization exceeded $1 trillion, successfully entering the "trillion dollar club".

Most of the credit may go to OpenAI's launch of ChatGPT, an industry event. Since its launch in November 2022, the number of registered users has soared to 100 million in three months.

These two sets of data unmistakably reveal that artificial intelligence is consuming computing power like never before. This also makes China's scholars and experts engaged in the computing industry feel pressure - in the face of the severe external situation, Chinese intelligent computing power services.

On May 29, the China Intelligent Computing Industry Alliance launched the 2023 China Computing Power Development Seminar, listing the above propositions as an important agenda of the conference, and inviting participating experts to provide suggestions.

Supercomputing Internet "assists" AI computing power services

"To make a big model, you must have a large computing power, the so-called 'vigorous miracle', the greater the computing power, the more likely it is to produce a miracle." Zheng Weimin, academician of the Chinese Academy of Engineering and professor of the Department of Computer Science of Tsinghua University, said in a forward-looking way that ChatGPT is based on the product of large-scale computing power infrastructure, where should China find enough powerful artificial intelligence computing power to train a more intelligent large model?

"This requires a new type of computing power infrastructure to specifically do corresponding computing power services." He said: "Of course, there are still many problems that need to be studied, including computing network connection, data transmission and so on. ”

The answer may be at hand. In April this year, China's computing circle ushered in a major event: the Ministry of Science and Technology launched the national supercomputing Internet deployment work. According to the deployment, the national supercomputing Internet connects many supercomputing centers across the country through the computing power network, operates the supercomputing center with Internet thinking, builds an integrated computing power service platform, realizes the overall scheduling of computing resources, lowers the threshold of supercomputing applications, and drives the development of computing technology to a higher level.

According to Cao Zhennan, deputy director of the National High Performance Computer Engineering Technology Research Center, at present, the national supercomputing Internet platform has entered the research and development stage, and will gradually implement the overall layout of advanced technology, model innovation, high-quality service and perfect ecology.

How much help will the supercomputing Internet platform bring to AI computing?

"Although people often divide computing power into supercomputing and artificial intelligence computing, there are many similarities between supercomputing and artificial intelligence computing in mode and application, and the difference is only in the different requirements for computing accuracy." Cao Zhennan said that at present, the demand for computing power in artificial intelligence has a large gap, and the accuracy requirements are not high in pursuit of computing performance, but if computing power becomes easy and cheap, "some AI applications may put forward some requirements for accuracy."

Cao Zhennan believes that the boundary between supercomputing and artificial intelligence computing is becoming more and more blurred. At present, AI computing power is unevenly distributed in different regions and industries, resulting in a huge AI computing power gap in many industries. One of the construction goals of the supercomputing Internet platform is to solve the problem of uneven distribution of computing power facilities and realize the interconnection and resource sharing of computing power resources.

At the seminar, Qian Depei, academician of the Chinese Academy of Sciences and leader of the supercomputing Internet overall expert group, was also among the discussion experts, and he had a deeper view of computing services.

"Our computing power infrastructure should try to provide not the 'cornerstone', but the software or services required by users, which is actually the essence of the computing power Internet, only in this way, people do not care what processor provides what computing power behind it." Qian Depei said that this will have an impact on the survival mode and service model of supercomputing centers or intelligent computing centers, and in the long run, this can truly penetrate computing power into all walks of life and become part of production and life, "Of course, there is still a long way to go."

Computing power services should be reasonably planned to avoid waste

In any case, with the change of artificial intelligence technology and applications, the demand for computing power in artificial intelligence applications is increasing, and the trend is unstoppable.

"We have a judgment that we expect to experience the wave of generative AI represented by ChatGPT, and artificial intelligence computing is accounting for an increasing proportion of data centers." Chen Wenguang, a professor of computer science at Tsinghua University, said that he observed that compared with traditional supercomputing centers, the utilization rate of intelligent computing centers is higher, "which also shows that AI is becoming more and more penetrating into all aspects of economy and society."

"We should embrace this trend." Lu Zhonghua, a researcher at the Computer Network Information Center of the Chinese Academy of Sciences, believes that now, all walks of life have different views and concepts on computing power infrastructure, computing power interconnection, supercomputing intelligent computing platform services, etc., but only one thing is the same, that is, people's demand for computing power is growing, and the demand for artificial intelligence applications in this regard is more prominent.

"How do you meet these computing power needs? I have three suggestions. Lu Zhonghua said, first of all, in the next 5~8 years, we must strive to make good use of the supercomputing center and intelligent computing center that have been built, "do not waste the existing computing resources"; Secondly, she said that the potential commercial market of generative artificial intelligence represented by ChatGPT should not be underestimated, and at this time, "Chinese must dare to use domestic AI chips to do ecological construction for China's own large model development"; Third, regarding the "construction of an integrated artificial intelligence computing power platform", we cannot "wait for what is needed", specific to the supply and demand of computing power of all parties, and we must hand over the market to the market.

However, for users who need to use artificial intelligence for research and development, there are still expectations for the layout of artificial intelligence computing power services at the national level.

"AI computing power service, of course, we hope to have a layout, but I think there must be both a layout and a division of labor." Chen Runsheng, an academician and bioinformatician of the Chinese Academy of Sciences, said bluntly at the seminar: "China should have a large model that can benchmark OpenAI, but at the same time, it must also have different artificial intelligence computing services." This requires a division of labor according to the capabilities of all parties, rather than many families doing the same thing, otherwise resources will be wasted. ”

It is important to build a domestic computing power network ecology

In the matter of worrying about the waste of resources, Zhang Yunquan, a researcher at the Institute of Computing Technology of the Chinese Academy of Sciences and secretary-general of the High Performance Committee of the Chinese Computer Society, observed that since the concept of computing power network was proposed, various paths and concepts have emerged on the interconnection of computing power from all parties, and there is no definite standard and route.

"I think one of the problems we are facing now is that we need to find a practical way from the national macro level to plan the interconnection of computing power." Zhang Yunquan said that this process should be fully demonstrated, "don't find the wrong way, otherwise it will be very troublesome, resulting in waste of resources and delays."

The construction of the national supercomputing Internet platform has taken the lead in providing computing services.

"I think that whether it is supercomputing or intelligent computing, there are many bottlenecks and problems in the service process, but the most difficult thing is the application problem." Cao Zhennan said that in his opinion, at the level of computing power, the country should develop software and hardware in a coordinated manner, and at the same time develop the domestic computing power network ecology.

"A very important purpose of building a supercomputing Internet is to build an ecosystem of China's supercomputing and China's computing industry, hoping that different entities will form standards through ecology and create a more reasonable market model." Cao Zhennan said that this is also an opportunity for China's computing service market to move towards a new stage of development under the background of artificial intelligence applications putting forward higher demand for computing power.