The real competition in the artificial intelligence industry has just begun

Author ｜ Sang Mingqiang

Similar to the consumer Internet in the past, in recent years, the artificial intelligence industry has a little bit a little bit a little bit more meaningful. On the one hand, the solutions of different scenarios are separated to make the threshold of standardization higher; on the other hand, Gartner mentioned machine learning, marginal AI, decision -making intelligence in the "2021 AI maturity curve" in "2021 AI maturity curve". Innovation will have a change in the market, but it is not easy to break through these technologies in the short term.

In this situation, the "turning point" and "how to break the situation" have become a key proposition for players to think. Although the uncertainty of the extension of the AI ​​format is intensifying, the digital economy and digitalization have become the industry consensus. From the rise of AI star companies to the implementation of new infrastructure, east counting, Positive feedback, then the latter is a industrial -grade framework.

For a long time, artificial intelligence has been regarded as an important tool for helping humans break through scientific and technological progress cognitive obstacles and cognitive boundaries. The "smart car" representing the "Yuan Universe", including Zuckerberg's bet, has set off an industry -level change wave. According to the three elements (data, algorithms, and computing power) of AI, if it was said that it was a period of prosperity of data and algorithms in the past, then now it is obviously the golden age of computing power.

According to the calculation of China Xintong Institute: In 2021, the scale of the core industry of the computing power has exceeded 1.5 trillion yuan, of which the cloud computing market has exceeded 300 billion yuan, the IDC (Internet data center) service market has exceeded 150 billion yuan, the size of the core industry of artificial intelligence is More than 400 billion yuan, the pillar role of the computing power industry became more and more prominent.

Generally speaking, computing power is computing power, which refers to the processing capacity of data. It is small to mobile notebooks, as large as supercomputers, and the computing power exists in various smart hardware devices. Each face recognition and message of each person during the security check Each voice conversion requires the computing power support of the hardware chip. It can be said that the size of the computing power represents the strength of digital information processing capabilities, and it is also a key indicator for measuring the gold content of an AI company.

In the past two years, the definition of great computing power has been opened to a certain extent. Many international technology companies include Google, Tesla, Facebook, etc. At the expense of huge sums of money to build an artificial intelligence computing center, many domestic governments and enterprises have also bowed into the bureau and build a new generation of artificial intelligence infrastructure. However, there are also opinions on this. At the 2022 China Computing Power Conference, Yang Fan, the founder of Shangtang Technology and President of the AI ​​Device Group, believes, "When we are talking about computing power, we must think about it. What value can we bring to us. "

From the choice of survival at the past business level, tired of catching up, to the underlying innovation and game of basic science, it is obvious that the domestic AI industry and players are standing on a new turning point in history. Essence

The golden age of 01 computing power has just begun

We often say that today's world has experienced a large change in a century, reflected in the digital economy, a new round of scientific and technological revolution and industrial changes that are driving with digital technology at the core of digital technology are profoundly affecting the path of economic growth and the direction of productive forces. This is a phenomenon that is worthy of attention and in -depth research. A few days ago, Ren Zhengfei once pointed out: The era of great power has arrived, we are turning, we must understand and participate in the changes in this era.

In the past years, the domestic computing power industry has developed very fast, and it has basically been consistent with the current digital development. According to a set of data given by Yu Xiaohui, the dean of the China Xintong Academy, the average growth rate of China's computing power industry in the past 5 years has exceeded 30%; unique, according to a study released by OpenAI before, the computing power used in AI training is every 3 every 3 computing capabilities used in AI training every 3 every 3 It will be turned over by 4 months. However, what people do not expect is that the computing power of the chip will reach the limit of a day, and I do not expect that the chip computing power limit will come so fast.

Earlier, researchers at the Massachusetts Institute of Technology had issued computing power warnings -deep learning is approaching the calculation limit, which is also the pain point of the current computing power industry, that is, the demand for AI algorithms and application scenarios in the industry is increasingly diverse. The collaboration of upstream software design and chip hardware architecture has a challenge. In popular terms, computing power is a new productive forces in the intelligent era. This is beyond doubt, but the problem appears on the dimension of "productivity", because what everyone does not know is how important the computing power is, how expensive it is, how expensive it is. Essence

As early as 1961, at the beginning of the definition of computing power, Professor John McCarthy had proposed that computing power should be used as water and electric resources. Similar to today's cloud computing It is a pity that to this day, this idea still faces many core technical challenges. For a simple example, the core of many AI chips at the moment is to use the multiplication array to achieve the most important convolutional computing in convolutional neural networks, but the large amount of operations of the Mac array will increase power consumption. Excellent performance and power consumption ratio has always been the key goal of AI chip research and development.

But this is not without solution. Some domestic AI players have tried to give their own problem -solving ideas, and even in some aspects in front of the world. Take the large device of Sensecore Shangtang AI as an example. It is a new type of artificial intelligence infrastructure created by Shangtang. It takes Shangtang AIDC (new artificial intelligence computing power center) as the base to organically integrates the deep learning platform and model layer organic integration organically While the computing power can be better decoupled and utilized, it has also achieved batch algorithm model production, deployment and iterative upgrades. In other words, the problem of large computing power is not only technical, but also a cost issue. Some people in the industry have made a more vivid metaphor: if the GPU is compared to a bus, the calculation task of AI is equivalent to the tour group. In the traditional computing power distribution mode, whether the calculation is large or small, it will occupy a hardware. The unit, like a small group traveling, has covered the entire bus, which is undoubtedly a waste of resources.

This is actually a key contradiction in the artificial intelligence industry today -technical costs and commercialization are not matched. If AI solves a problem, the cost and price it takes is far greater than the benefits of solving the incident. Then this matter should be considered seriously, because in the final analysis, the computing power is good, and the algorithm is to be able to be able to be able to be able to be able to be able to be able to be able to be To better solve the problem and achieve relatively positive commercialization results.

02 Thinking: Scene construction and quadrant focus

As mentioned earlier, a key proposition of large computing power now is how to solve the needs of long tails, but in the actual application scenario, the chain from computing power to terminal demand is long, including computing power hardware, computing power scheduling management Software, capacity platform software, scene solution services, each key link is a relationship between upstream and downstream, the lack of any link will affect the effect of the final computing power landing application.

This is the highest challenge put forward in the era of intelligence, because the essence of intelligence is the in -depth processing and refining of data, relying on computing power and algorithms to train data into models, and then realize general artificial intelligence, explain artificial intelligence, and even categories. Intelligent modeling, including brain intelligence, eventually forms a smart era represented by large models. This is also the three elements of the AI ​​era in the Xu Xinkou of the "Queen of the Venture Capital": a good solution must achieve computing power, algorithms and Data quadrant focus.

We still take the more typical Shangtang Technology as an example. It is very determined to invest in large computing power. It is reported that in Shanghai Lingang, the artificial intelligence computing center built by Shangtang has begun this year Power is also one of the largest artificial intelligence computing centers in Asia. From the perspective of Shangtang, the current task is to figure out the space, data, and algorithm demand space for the future of AI, but also to know your positioning: when you have a better AI infrastructure, you can in the end What value to create users and customers in various industries.

At the 2022 China Computing Power Conference Business Tang Artificial Intelligence Infrastructure Innovation Forum, the Northern AI computing power innovation center led by Zibo Municipal Government officially announced the launch

According to Xu Li, the co -founder and CEO of Shangtang Technology, "the essence of AI devices is to allow AI to get rid of human dense state." He believes that the reason why AI is dense is because the production efficiency is not high, and the production efficiency is improved. The key is that the cost of production factors is low enough. According to Xu Li's "machine conjecture", the way of human brain may never keep up with the speed of the universe expansion. If you want to really solve the truth of the universe, you can try to use larger data. A stronger computing power may also collide with more expected results.

Taking smart cities as an example, today's entire city -level service and comprehensive governance can be completed through a lot of video processing and analysis. Of course, it also means that better AI infrastructure and better computing power support are required Essence At present, AIDC has also implemented many industrial applications in Shandong, including smart parks, smart remote sensing, smart medical care, smart education, etc. Based on the city -level industrial computing power base of the Shangtang AI computing power innovation center, the "smart capital" is allowed Zibo's new business card also slowly surfaced.

To some extent, the "Shang Tang Mode" is not complicated. In fact, it is the way to build an industrial ecology and logic, that is, the AI ​​computing power innovation center is the fulcrum, and the enterprises in the central city and the surrounding areas are available for AI services, thereby forming a divergent Although the industrial chain structure needs to be verified in many aspects, it also has a feasible sample for the domestic AI industry.

This is easy to remind me of the book "Fusion of Cloud Network: Digital Information Infrastructure in the Era of Computing Power". The author Li Zhengmao has made a more vivid explanation in the book: if the Internet is an intelligent era 1.0, computing power network It is the Intelligent Age 2.0, and the model network is the intelligent era 3.0. From the perspective of information, the telecommunications network solves the data networking, the Internet solves the information network. Complete the infrastructure of information.

03 Summary

There is such a saying that when electricity becomes an infrastructure or the main driving force for the industrial era, humans can really enter the era of electric vapor. It is now everywhere, but it is not enough. We must admit the current artificial intelligence. At best, it can only be said to be the first shot before the golden age of computing power.

As the deep learning network model is becoming more and more universal, the parameter scale is getting larger and larger. Requirements for the increase in algorithm accuracy. In addition, there are more and more long -tail segments in the industry. Although these applications are low, once irreversible effects will occur, the industry is huge. "There are still a lot of gold waiting to be excavated. Back to the topic of discussion at the beginning, how to understand the key indicators to test artificial intelligence. In fact, it contains 2 layers of meaning: 1. The "scene upgrade" logic is "scene upgrade", and the current logic It is not the "subversion" logic in the Internet mouth, so you need to take root in the bottom scene needs to achieve "adapting to local conditions"; Click, just like the "last mile" problem in the logistics field, it is a hurdle that must be across.

- END -