At the end of May 2021, the National Development and Reform Commission, the Central Network Information Office, the Ministry of Industry and Information Technology and the National Energy Administration jointly issued the Implementation Plan for the Computational Power Hub of the National Integrated Big Data Center Collaborative Innovation System. In this plan, the overall and coordinated development of computational power resources was set at the macro level, and a new vision of the future network emerged;
Just four months later, at the Huawei All-Connection 2021 just held, the China Institute of Science and Technology Information, AITISA (New Generation AI Industry Technology Innovation Strategic Alliance) and Pengcheng Laboratory jointly released the “White Paper 2.0 on the Development of AI Computing Centers – From AI Computing Centers to AI Computing Networks”, which completed the macro vision and road map design of the new vision of the future network, Start to take steps and put them into practice.
AI computing network: the killer of AI industry development in China alone
The AI computing network is becoming one of the biggest waves in the “new infrastructure” wave, driving the development of China’s AI industry into an epic acceleration process, bringing the ultimate goal of AI development, the intelligent society, closer. In this process, China’s unique advantages in developing AI industry have also been fully demonstrated.
As for the answer of computing power, “dimension” is gradually increased. The emergence of AI computing power network is inevitable
To understand what AI computing power network is and why it appears, we should start from the process of the industry’s continuous demand for computing power.
With the accelerated landing of AI scenario-based applications, among the AI troika, innovative algorithms are emerging and data are becoming more and more abundant, while the challenge of computing power is becoming increasingly severe – the supply of computing power at the upstream of the AI “industrial chain” is beginning to be insufficient, forcing the advancement of various forms of computing power supply, and “upgrading” step by step.
Stage 1: Basic software and hardware innovation
At the beginning, the basic software and hardware for AI computing developed rapidly, including the launch of specialized chip products (such as NPU), or further, heterogeneous computing based on the innovation of underlying computing architecture (such as Ascension AI), which provides AI with more computing power locally or in the cloud than in the past.
Phase II: construction of artificial intelligence computing center
Then, with the further development of AI industry, in the face of stronger demand for computing power, even with the innovation of basic software and hardware, when a single enterprise deploys or purchases computing power, it will also face the problem of shortage of computing power resources or high price.
Therefore, in those places with AI industrial clusters, artificial intelligence computing centers led by the government began to appear, providing enterprises with relatively more reasonable computing power prices and more flexible supply modes through an intensive way. The most typical examples are Shenzhen Pengcheng Cloud Brain II and Wuhan Artificial Intelligence Computing Center, which have been launched. With the support of Shengteng AI, their computing power resources are running at full capacity, which shows the strong demand.
At the same time, there is the industrial layout of AI computing center.
Unlike the same intensive deployment of supercomputing centers, which only serve scientific research well, AI computing centers have an industrial mission. To help the development of smart cities, smart finance, smart manufacturing, smart transportation and other industries, pure computing power also needs to coordinate with industrial development. Therefore, local governments have formed a pattern of one center and four platforms to build AI computing centers: relying on AI computing centers, It has built a public computing power service platform (inclusive computing power), an industrial application innovation incubation platform (to create an AI application demonstration benchmark), an industrial aggregation development platform (to promote the intensive development of AI industry), and a scientific research innovation and talent training platform.
Stage 3: AI computing network
After the continuous development of computing infrastructure including AI computing center, new problems emerged:
AI computing power in different regions also has its own peaks and valleys, and the AI big model developed on the basis of more abundant computing power has a phased and high-density use of computing power. Naturally, people in the industry began to think about the allocation of computing power resources across regions, connecting the computing power network between AI computing centers in different regions, and realizing resource sharing and collaborative scheduling.
Therefore, the artificial intelligence computing power network came into being. It uses new network technology to connect the nodes of the artificial intelligence computing centers distributed around the region, sense, distribute and dispatch the artificial intelligence computing power in the region, and dynamically allocate the computing power according to the situation of the computing power resources in each center and the demand of each region.
In view of the current development of AI computing centers and China’s regional economic situation, the future AI computing network may refer to the “three plus one” policy proposed by relevant experts in terms of the specific interconnection mode: relying on the interconnection construction of the Yangtze River Delta, Guangdong, Hong Kong, Macao Bay Area, Beijing-Tianjin-Hebei, Chengdu-Chongqing Economic Circle and other regions, step by step.
It is worth mentioning that in addition to computing power, data and algorithm resources are also gathered and integrated across the country through the AI computing power network. The value concept of “one network, three convergence” is also proposed in the white paper. In short:
One is the artificial intelligence computing network;
Convergence of computing power, i.e. high-speed interconnection of nodes, unified management and operation and maintenance;
Data aggregation is to realize the safe and open public data resources of different nodes and build high-quality AI public data sets;
Ecological convergence, that is, the opening of large model capabilities at different nodes and the sharing of application innovation achievements, aims to strengthen cross-regional scientific research and industrial cooperation.
It can be seen that from the innovation of basic software and hardware to the artificial intelligence computing power network, the supply of computing power resources is becoming more and more sufficient, and the synergy of computing power with algorithms and data, and the integration of computing power with industry are also becoming closer and closer. Of course, there is no strict time sequence for these three stages, and they are also being carried out at the same time.
Why can only China build a good AI computing network?
The emergence of AI computing network is an inevitable result of the development of industrial demand. However, it can only happen in China. It is determined by China’s unique AI technology and industrial development reality, and it is also the unique advantage of China.
The reasons include at least four aspects:
1. Background of synchronous development of new infrastructure
On the one hand, the artificial intelligence computing network itself exists as a new type of infrastructure. On the other hand, it needs to be supported by the construction of a new infrastructure at the lower level.
For example, the overall planning of the computing power of different nodes requires a communication network that can support the transmission and access of massive data. According to the plan, in the next decade of its implementation, broadband will achieve the leap from gigabit to 100G, while IP and other resources will support a hundred-fold increase in capacity. In addition, a strong edge computing software and hardware infrastructure is also needed to realize the center+edge distributed computing model to achieve 100 times capacity growth.
From AI computing power network to industrial application, a lot of new infrastructure is also needed in this process. From the production of computing power to the final use and good use of computing power, there is no good new infrastructure foundation can be implemented.
While China’s new infrastructure has been in vigorous progress from the macro policy to the specific implementation. The artificial intelligence infrastructure has been clearly identified as the core task of the new infrastructure. At the policy level, it is also the infrastructure supporting the self-reliance of science and technology and the development of the digital economy. 5G, wifi6, IPv6 and other technologies are being widely popularized, which is unique in the world.
2. Rapid development of “node” itself
From the formation of AI computing network, we can find that this is a macro-level overall planning, and its implementation process is not a construction process from having nothing to having everything. The “nodes” of AI computing center become the core resources. If there is not a large number of “nodes” or can not promote the rapid construction and landing of “nodes”, the AI computing network can only be a castle in the air.
This is why there are many discussions about AI computing networks, but few can be put into action.
In China, the AI computing center has long been the landing entity of the “building efficient computing infrastructure” in the State Council’s “New Generation AI Development Plan”. So far, under the leadership of the local government and the participation of Huawei and other scientific and technological enterprises, artificial intelligence computing centers have been built and put into operation in Shenzhen, Wuhan, Xi’an and other cities, and artificial intelligence computing centers in Chengdu, Central Plains and other places are under construction. In addition, the construction of artificial intelligence computing centers in Beijing, South Beijing and other places is also under planning.
These intensive nodes, which only appear in China, promote the rapid landing of AI computing network.
3. Demand convergence of industrial cluster development
Without the support of huge demand, or the existing demand in the market cannot be connected to effectively “supply” computing power, the AI computing network cannot be implemented, or the implementation process itself is a waste of resources and energy.
At this time, the value of the unique industrial cluster development path in the process of China’s social and economic development emerged.
On the one hand, because the development of AI industry in all regions is carried out in a clustered way, the demand is huge and concentrated, and the effective demand of the region can be gathered together to realize the full connection between supply and demand;
On the other hand, because of the huge demand of industrial clusters, there will generally be computing power peaks and valleys of their own. Under the artificial intelligence computing network, effective computing power supply demand can always find appropriate users locally or across regions.
This enables the AI computing network to achieve “make the best use of everything”, maximize the value of landing and stay away from “dragon killing”, and form a healthy development.
Finally, the efficient operation of AI computing network will promote the rapid development of various industrial clusters in the country, and a positive feedback cycle will be formed.
4. Scientific and technological enterprises with internal drive to break through are making efforts
Looking further at the details of the implementation, AI computing network needs a lot of technical innovation to support it. For example, only in terms of the integration standards of computing network, it is necessary to complete the construction of computing network architecture and interface, application and computing power perception research, computing power demand quantification and modeling research, computing network resource trust and collaboration and other standards.
If these down-to-earth technologies do not break through, they will become the weakness of the barrel, seriously affecting the computing power of the AI computing network and the industrial promotion effect.
However, since the AI computing network itself is a new thing, many matching technologies are also new, and it is difficult to find referential objects around the world. At this time, Chinese science and technology enterprises with internal drive for independent innovation under special and complex circumstances will be more willing to invest in these technologies, which is an important opportunity to catch up with the world and establish a voice in technology.
It is not to say that foreign science and technology enterprises can’t make these technologies, but Chinese science and technology enterprises, represented by Fang Hua, the solution provider of the artificial intelligence computing center, are more motivated on the one hand, and on the other hand, in a new field, based on the existing new infrastructure experience of computing power (such as Ascension Full Stack AI), will be better at technical understanding.
From technological innovation, industrial promotion to technological highland, the value of AI computing network is highlighted
With a sharp sword in hand, there are too many answers to explain what the AI computing network can do, but at least three aspects are of certain value.
First of all, it is strong support for breakthrough technological innovation.
In this respect, multimodal large model is typical. As an important technological innovation of strong AI and general AI for the future, multimodal large model has been put forward in the field of AI for many years, and there are many technological breakthroughs in the industry. However, the multimodal large model has gone further down, and the demand for computing power has risen geometrically, and the general computing infrastructure will soon be unable to be competent.
From the single training of image, text and voice to the transition of dual-mode and three-mode, AI can flexibly cope with the transformation of different modes, as natural as the interaction between people and the world. In this regard, it will have obvious advantages to be driven by the AI computing network.
Similarly, there are many technological innovations that need to be supported by explosive growth of computing power, which will be rapidly promoted under the support of AI computing power network.
Then, it is to fully meet the needs of the whole chain of industrial development.
As we all know, the extensive application market is an important advantage for China’s development of all technologies. Putting technology on the market generates value and feeds back technology research and development, which has become the standard model for many industrial technological progress.
However, in AI, upstream supply and downstream demand become equally important. The emergence of AI computing power network means the development of China’s AI industry and technology. On the one hand, it still has a huge scene application space to achieve business value drive, and on the other hand, it has a leading global computing power supply capacity, with more obvious comparative advantages, promoting the prosperity and acceleration of the AI industry.
Finally, it is to build and form a real technological highland.
In the past, the development of AI has seen rapid progress in computing power, algorithms and data, but in the final analysis, there are not many breakthroughs. Most of them are continuously optimized and improved under the existing framework. For example, if the hardware fails, the chip performance will be improved, and if it fails, it will be stacked to win by quantity.
The emergence of AI computing power network may have a similar breakthrough significance with the creation of basic models at the algorithm level (for example, the BERT model in the NLP field). It has directly promoted the development dimension of AI in the way of assimilation. After the AI computing center, it has further broken through the common form of the old computing power supply, allowing AI to truly get rid of the shackles of computing power and meet new thinking and practice. This also shows that under the power of integration, China is occupying a new technological highland belonging to artificial intelligence.
In short, in the past many years, China has continued to lead the world in Internet technology, artificial intelligence and other fields. Now, this leadership is continuing and deepening. With the support of artificial intelligence computing network, China is expected to be the first to enter the intelligent society, and truly gain the voice of technology and industry chain that overlooks the world.
