Zhi DongXi (public number: zhidxcom)
Author | Yang Chang
Edit | Desert Shadow
2021 is the opening year of the "14th Five-Year Plan", and under the wave of large infrastructure, artificial intelligence (AI) infrastructure has also ushered in unprecedented development and technological product iteration.
As AI is more widely implemented in thousands of industries, the demand for ARTIFICIAL computing power in regions and industries is also increasing, and AI-oriented computing infrastructure has become the focus of new infrastructure.
The intelligent computing infrastructure represented by the intelligent computing center (hereinafter referred to as the intelligent computing center) is a new type of public computing power infrastructure that provides the required computing power services, data services and algorithm services for AI applications, which is composed of AI chips and computing units and other equipment, gathering massive data, and enterprises and research institutions can rely on the powerful computing power provided by the intelligent computing center to drive AI models for data deep processing and realize AI application innovation.
The purpose of the construction of the intelligent computing center is to promote the industrialization of AI and the aiization of industry.
The smart computing center is different from the cloud data center, the cloud data center is the use of virtualization and other cloud computing technologies, providing still the traditional data center business and a variety of new network applications, and the smart computing center is the virtualization of the data center, may be part of the cloud data center in the combination of the combination of computing power, may also be multiple cloud data center collaboration to provide computing power.
However, compared with traditional data centers, the construction history of intelligent computing centers is shorter and there are fewer standards in the industry. As a new type of infrastructure, what kind of computing power does the intelligent computing center need to provide to society? Under the dual-carbon policy, how can the Intelligent Computing Center achieve greener? In the face of the huge AI intelligent computing industry chain, how to coordinate upstream and downstream co-construction? How can the intelligent computing center built with huge funds better support local economic and social development?
In view of the key and important issues in the construction and operation of these intelligent computing centers, Ren Jingxuan, senior vice president of Shuguang, conducted in-depth exchanges with zhidong and other media and shared the shuguang plan.
First, the intelligent computing center has become a hot spot for new infrastructure, and four major problems need to be solved
It is not easy to build and build a smart computing infrastructure, Ren Jingxuan said, like the smart computing center and other smart computing infrastructure, its development is still in its infancy, construction, will face some initial difficulties.
▲Ren Jingxuan, senior vice president of Sugon
The purpose of the construction of the intelligent computing center is to serve the local economic development, that is to say, the intelligent computing center is first of all a public computing power platform, not only for an enterprise or an industry, but also needs to serve as many local enterprises or types of industries as possible.
The intelligent computing center cannot be an "island", its applicability cannot be limited, and it cannot build an intelligent computing infrastructure with low utilization rate or can only support a small number of applications.
This is the first problem that the construction of smart computing infrastructure needs to face, how to meet the needs of regional and local enterprises, and even provide a wider range of services, so that the construction of smart computing centers is more meaningful.
The second problem in the construction of the intelligent computing center is that with the "double carbon (carbon peak and carbon neutrality)" goal, the double carbon planning in various places has been released, how to reduce carbon emissions has become a hot topic, and how to reduce the energy consumption and reduce carbon emissions of such energy-consuming households as the intelligent computing center is a problem that the builder must consider.
The problem that needs to be solved in the construction of the third intelligent computing center is how to coordinate the upstream and downstream forces of the AI industry chain to jointly build a smart computing center.
From the lowest chip, to the middle layer of the framework, model, algorithm, and then to the top level of the application, the AI industry chain is quite long, each layer has a number of subdivision tracks and a large number of players, how to integrate the strength of each layer at this stage, the intelligent computing center to build a good, is a very realistic and meaningful topic.
The fourth problem that needs to be faced in the construction of the intelligent computing center is how to do a good job in industrial empowerment.
The purpose of the construction of the intelligent computing center is to promote the development of the local industrial economy, and in the current era of AI empowering thousands of industries, investors will face various challenges in the process of using the intelligent computing center to accelerate the industrialization of AI and the AI of the industry.
Second, when AI is landed, how to build, manage and use the intelligent computing center?
In the face of the above-mentioned difficulties in the construction and operation of the intelligent computing infrastructure or the intelligent computing center, Shuguang proposed the construction of "5A-level" intelligent computing infrastructure through its own practice and summary, hoping to promote the development of industrial ecology and coordinated development, stimulate industrial vitality and overall prosperity, and provide an important support carrier and driving force for the integrated development of the green economy and the digital economy.
▲Shuguang "5A-level" intelligent computing infrastructure
Shuguang constantly emphasizes the "5A-level" intelligent computing infrastructure, which specifically includes the capabilities of openness, integration, green, inclusiveness and service.
Openness is to build a diversified integrated infrastructure of the intelligent computing center in an ecological inclusive mode; integration is to carry multiple types of chips to provide multiple computing power; green is to use advanced phase-change full immersion liquid cooling technology to reduce energy consumption by up to 30%; Universal is to reduce the comprehensive construction cost to 30% lower than the existing market price through a variety of strategies; the service is to provide a full-life cycle of intelligent computing infrastructure services, from construction to operation, continuous service.
1. Openness: With the purpose of diversified collaboration, the pursuit of industrial chain co-construction
At present, the primary problem facing the intelligent computing center is the problem of openness, and the increase in the number of AI applications and the speed of iteration challenge the vitality of the intelligent computing center.
As a new type of infrastructure, the Center for Intelligent Computing cannot only solve problems in one field, such as materials or genes.
At the same time, the construction of the intelligent computing center should also avoid this situation: after spending a huge amount of money to build the intelligent computing center, for example, after one or two years, with the iteration, the function of the intelligent computing center will be greatly reduced.
There are two ideas in the industry for the construction of intelligent computing centers, one is a vertical integration integration model, each layer is independently realized by the builder, so the advantage is to avoid the adaptation process between different technical routes.
The second approach is a multi-collaboration model, which may be slightly more difficult to adapt, but it will be more friendly to developers and more beneficial to industrial development.
Shuguang adheres to the multi-collaboration model, Ren Jingxuan said, Shuguang construction of "5A-level" intelligent computing infrastructure, that is, through the full openness and compatibility of chips, algorithms, frameworks, models, to build a diversified integrated infrastructure, to achieve the maximum ease of use of the computing base, reduce migration costs, bring a better environment to future developers, and then drive the healthy development of the entire industrial chain through the construction of intelligent computing centers.
2. Convergence: Through the distributed heterogeneous parallel architecture, the supply of diversified computing power is realized
With the continuous development of AI technology and the improvement of the understanding and needs of various industries for intelligence and digital transformation, from the perspective of the industry horizontally, there are more and more fields empowered by AI, from smart cities, to industrial manufacturing, to agricultural production, scientific computing, etc.; from the vertical perspective of the industry, AI applications are gradually developing in depth in vertical fields to solve key problems within the industry.
AI applications are diverse, the demand for computing power is also diverse, both applications require high-precision computing power, there are also applications that require low-precision computing power, hybrid precision, multiple computing power has become an inevitable trend in the development of intelligent computing infrastructure.
Shuguang's "5A-level" intelligent computing infrastructure is equipped with multiple types of chips through a distributed heterogeneous parallel architecture, which can cover the accuracy of full computing power, thereby realizing diversified computing power supply to meet the needs of different AI application scenarios and multiple users.
3, green: the use of phase change type full immersion liquid cooling technology, the measured PUE value up to 1.04
In addition to the increasing application of AI, the parameters of AI models and the computing power volume of the intelligent computing center are also increasing year by year.
Especially now that there are more and more large models, taking GTP-3 as an example, the relevant data shows that if the entire program of GPT-3 is run again, the carbon emissions generated by the model are equivalent to a car driving 700,000 kilometers, which is about twice the distance from the Earth to the moon.
Under the dual carbon target, the form of energy conservation and emission reduction of the Intelligent Computing Center is quite severe.
Shuguang's "5A-level" intelligent computing infrastructure adopts a variety of ways to reduce the energy consumption of the intelligent computing center, of which liquid cooling technology is the killer.
The PUE (Power Usage Efficiency Value) value is an important indicator for evaluating the carbon emissions of intelligent computing centers and data centers. In the future, data centers or intelligent computing centers with values above 1.5 will be phased out.
Shuguang builds a green and low-carbon "5A-level" intelligent computing infrastructure through multi-level technological innovation such as clean energy and energy saving of chips, equipment and platforms.
In addition, based on Shuguang's innovative immersion phase change liquid cooling technology, the PUE value of the Intelligent Computing Center can reach 1.04 to 1.05.
Dawn is also working on technologies such as waste heat reuse, hoping to reduce the PUE value of the data center to a lower level.
4. Universal benefit: optimize the cost of computing power, so that the comprehensive construction cost is 30% lower than the existing price in the market.
At present, there are two characteristics of the price of the intelligent computing center, one is expensive, and the other is chaotic.
Expensive because the intelligent computing center as a new type of computing power infrastructure, if the construction cost is converted into the unit price of computing power, the cost of computing power is very high. The chaos is because this is an emerging market, and the price of computing power is related to many factors such as the region where the smart computing center is located.
In the long run, there must be a very inclusive computing power to provide AI users and developers, so that the AI industry can continue to develop and reduce the investment and construction pressure of local governments.
Through a variety of strategies, Shuguang's "5A-level" intelligent computing infrastructure continues to optimize the cost of computing power, achieving a comprehensive construction cost that is 30% lower than the existing market price.
5. Service: Adopt an integrated service model to provide post-completion operation services
At present, most of the intelligent computing centers adopt the model of separation of construction and transportation, and there is no more mature business model in terms of operation, and the operating entities are generally in a vague area, and there may be a lack of operation.
However, for local governments that want to build smart computing centers to serve regional economies, the operation of smart computing centers is just a need and a problem that must be solved.
Shuguang's "5A-level" intelligent computing infrastructure adopts an integrated service model to achieve complete supporting services from planning and design, center construction, and platform operation, and runs through the whole life cycle of the intelligent computing infrastructure.
Third, can be landscape-level delivery, Dawn's intelligent computing productization landing road
According to our understanding, Shuguang's "5A-level" intelligent computing infrastructure can achieve ultra-high-density rapid delivery and landscape-level delivery, which refers to integration with the construction of the park and becomes a landmark of the city.
Specifically, in terms of ultra-high density, Sugon's "5A-level" intelligent computing infrastructure can achieve three-dimensional expansion, save 90% of the computer room area, and reduce the environmental requirements and cost of the computer room; in terms of rapid delivery, Sugon's "5A-level" intelligent computing infrastructure can achieve rapid deployment of infrastructure, computer room and equipment integration in 3-6 months.
The construction of Shuguang's "5A-level" intelligent computing infrastructure capabilities is inseparable from Shuguang's technical precipitation, advantage accumulation and experience prediction in the field of computing.
In July this year, the Chinese Academy of Sciences Artificial Intelligence Industry-University-Research Innovation Alliance, which is integrated by Shuguang and other relevant enterprises of the Chinese Academy of Sciences, released a new generation of artificial intelligence computing platform, which provides a reference basis and construction standards for the construction of intelligent computing centers from the dimensions of infrastructure, construction path, price model and so on.
In order to better land and productize the smart computing platform framework, Shuguang has made some attempts.
In September this year, Dawn announced the concept of "walking with the forwarders" and built a "5A-level" intelligent computing capability.
Recently, Shuguang has refined the "5A level" of smart computing infrastructure again, and accelerated the construction of smart computing infrastructure with five capabilities: openness, integration, green, inclusiveness and service.
Conclusion: The construction of the intelligent computing center is still in its infancy
Shuguang is the leading player in the field of core infrastructure in China, and has carried out many practices in the field of computing and intelligent computing, and proposed a higher-level concept of intelligent computing center construction.
Shuguang believes that the construction of the intelligent computing center is still in a relatively early stage, and the construction of the intelligent computing infrastructure should be carried out from the five aspects of openness, integration, green, inclusiveness and service, so as to improve the operational efficiency of the intelligent computing center.
With the continuous development of the AI industry and industrial AI, the construction of intelligent computing infrastructure continues to rise.
Compared with the traditional data center, the intelligent computing center may have a more obvious role in the development of the local industrial economy, and how to build, manage and use the intelligent computing center is a proposition of continuous iterative update.