With the continuous popularity of the concept of ChatGPT, the integration process of AI technology and thousands of industries has accelerated, and more diverse and complex application scenarios have emerged, which has put forward higher requirements for computing services. At the recent 2023 China Computing Power Development Seminar, many authoritative experts and scholars in the industry conducted in-depth exchanges and discussions on the opportunities and challenges of computing power under ChatGPT from the perspectives of technology and ecology, determined the development trend of the liquidation power industry, and gave countermeasures on the current artificial intelligence infrastructure and computing power service construction.
The key to the construction of computing power infrastructure lies in the combination of general and specialized courses.
As an important engine of computing power supply, artificial intelligence computing power infrastructure has ushered in a "big test" in terms of construction layout. According to incomplete statistics, more than 30 cities in China are building or proposing to build intelligent computing centers, which basically adopt the mode of "government-led, enterprise-contracted and joint operation". That is, it will be funded by local finance in a unified way, and will provide public computing services for all walks of life after completion. In this mode, the intelligent computing center is located in public service facilities. First, it must meet a wide range of application scenarios, be universal and realize universality. Secondly, it should also be able to support some personalized application scenarios that require high calculation accuracy and efficiency, and be efficient and dedicated.
Many experts at the meeting pointed out that the combination of "general specialization" will become the key to the construction of artificial intelligence computing infrastructure. Zhang Yunquan, a researcher at the Institute of Computing, Chinese Academy of Sciences, said that the emergence of ChatGPT realized the emergence of cognitive intelligence for the first time. At the same time, it also made the computing industry face new challenges such as rising demand, diversified development, prominent energy consumption problems and high use threshold. Computational power integration, factor coordination and ecological cooperation are the keys to the development of artificial intelligence computing infrastructure in the changing situation.
The most important thing for scientific computing is to improve its credibility, and so is artificial intelligence. Yuan Guoxing, a researcher at Beijing Institute of Applied Physics and Computational Mathematics, pointed out that to make the model more credible, it is necessary to continuously improve the accuracy of the model. Now, in order to improve the accuracy, the model is getting bigger and more complex, and the data increment is getting bigger and bigger. However, beyond a certain range, the cost of building a general model is too high, which requires a special model to solve the problem, so we should study this problem by classification.
Chen Runsheng, an academician of China Academy of Sciences, suggested that using large-scale academic infrastructure to build professional models can not only achieve high accuracy but also achieve low energy consumption. He believes that artificial intelligence computing services should have a layout and a division of labor. Don’t rush into it, don’t give up halfway, and avoid wasting resources.
Qian Yupei, an academician of the China Academy of Sciences, said that the question of generality or speciality has not only appeared in the field of large models, but has actually existed since the beginning of calculation. Heterogeneity is a trend, and different things should be done with different efficient tools. However, in the modern industrialized system, we should consider the cost, the performance is high and the energy consumption is low, but the design cost and manufacturing cost are too high. To compromise, we should not only achieve high performance and low operating cost, but also have low design cost and manufacturing cost, so that the whole system cost is low. With the tolerance of design cost and manufacturing cost, we should support different calculations with more efficient heterogeneous structures and components as much as possible.
Liao Xiangke, an academician of China Academy of Engineering, believes that the general model can serve all walks of life, and we need a general big model that can benchmark ChatGPT, and all walks of life can fine-tune and reason on the basis of the general big model according to the actual needs of the industry, and customize the special model of the industry.
Experts believe that on the one hand, the artificial intelligence computing infrastructure should have full-precision computing power and become a "generalist" of computing power, so that users can allocate computing resources according to practical application scenarios such as AI for Science and AI for industries, including general computing power, special computing power and even high-performance computing power to support their own business development. On the other hand, the artificial intelligence computing infrastructure of "combination of general and specialized" is a comprehensive scheme, which tests the openness of the underlying architecture. It is necessary not only for different technical routes to blossom, but also for adaptation and compatibility to achieve the same goal, so as to form an all-inclusive overall architecture, which will lower the threshold of application migration and promote the development of industrial ecology.
Computing network should connect resources and ecology.
The development of computing power network has also become a topic of great concern at present. The big model craze has brought huge demand for computing power, which has led to the uneven distribution of computing power between industries and regions. For the construction of computing power network, it is only the initial stage to unify the management and scheduling of various types of computing centers all over the country and realize the integration of computing power, storage, network and data. The more important construction link lies in the deep connection of the whole industrial ecology, that is, connecting people, connecting applications and services.
Zheng Weimin, an academician of China Academy of Engineering, said frankly: "Many institutions and enterprises in China are making big models, and each participant needs a lot of computing power support, which may cause a waste of computing power resources, which can be combined to form a big computing power, share it as a model, and make this model a new infrastructure. But there are still many problems to be solved to achieve this goal, such as how to connect different models, how to realize transmission, and how to meet the requirements for computing power.
"In terms of computing power interconnection, China currently has different schemes such as intelligent computing network and super-computing interconnection. Among them, the supercomputing Internet is to operate the supercomputing center with the thinking of the Internet, and connect the capabilities and resources of computing power supply, application development, operational services, users and other parties in the industrial ecology to build an integrated supercomputing network and service platform. In the tide of computing network construction, which road China will take in the end, how to connect our country’s computing resources, make good use of them, play a role and lower its threshold are all topics that need our discussion. " Zhang Yunquan said.
Supercomputing Internet, as an important form of computing power network, explains the essence of computing power network construction from the practical level. The first is the narrow sense of interconnection, that is, at the physical level, connecting computing centers of different architectures, building an infrastructure that can be used uniformly and serve the outside world, and realizing the scheduling and sharing of resources. Secondly, it is interconnected in a broad sense, that is, at the ecological level, the supercomputing center is operated with the thinking of the Internet, and based on the deep integration of resources such as computing, software and application solutions, an innovative platform led by application services is established, and the upstream and downstream are closely linked through the market-oriented operation and service system, so that the supply and demand sides can quickly connect and quickly find the resources they need.
"From the perspective of supercomputing the Internet, we hope not only to connect machines, but also to connect people, equipment and applications. The core is to be a supercomputing platform, and to make supercomputing and intelligent computing easy to use." Cao Zhennan, deputy director of the National High Performance Computer Engineering Technology Research Center, said.
Through the dual interconnection of physical and ecological aspects, computing network can further enable computing services, rationally allocate, integrate and release computing power, and lower the application threshold. Let computing power resources change from unattainable technology to universal and inclusive services, and support the development of major national scientific research projects, people’s livelihood and thousands of industries.
Cao Zhennan believes that there are still many bottlenecks and problems in the service process, whether it is super-calculation or intelligent calculation, and the most difficult one is the application problem. He suggested strengthening the coordinated development of software and hardware and attaching importance to the ecological construction of computing network.
Shan Zhiguang, director of the Information and Industry Development Department of the National Information Center, said that at present, algorithms and models are the places where there will be opportunities to make efforts in the future. ChatGPT is driven by centralized computing power. In the future, we may be able to take a different route, such as taking the next generation distributed route, and we can better connect some domestic computing infrastructures through more advanced interconnection technologies to tap stronger computing power.
"We still have to make a basic discussion from the evolution law of basic computing mode and the development law of artificial intelligence, so as to avoid rushing headlong into the crowd or following the trend at the phenomenal level, otherwise we can only run behind others forever. We must know that the computer field is competing for generations, and it is chasing further and further." Shan Zhiguang said.
"I think the concept of universality is changing." Tran van quang, a professor at Tsinghua University, said that the proportion of artificial intelligence computing forms in the whole data center will increase. However, at this stage, the AI Computing Center is still in a very early stage of development, and multi-party exploration should be encouraged at this stage.
Lu Zhonghua, a researcher in computer network information center, said that with the development of AI, the demand for computing power of AI services is increasing, and the trend is unstoppable. We should meet this trend. How to meet everyone’s increasing demand for computing power? In 5-8 years, we should try our best to make good use of the supercomputing center and intelligent computing center that have been built, so as not to waste the center resources that have been put into construction. We can take the lead in trying to develop large-scale model applications in some fields and encourage ecological construction. In addition, to build an artificial intelligence infrastructure system, we should not overemphasize the integrated layout, or we should support a hundred flowers under the guidance of national policies and hand over the rest to the market.
The Secretary-General of China Intelligent Computing Industry Alliance said quietly that the era of new computing power has arrived, and computing power will be the new kinetic energy and engine that will drive the digital economy forward in the future. At the same time, computing power is becoming a key factor affecting the country’s comprehensive governance and international discourse power. The core competitiveness of countries is focusing on computing power represented by computing speed, computing methods, communication ability and storage capacity. Whoever can master advanced computing power in the future will be able to grasp the initiative of development.