Title map 丨Visual China
In 1973, the Ministry of Electronics Industry of China mobilized universities and related R&D institutions to jointly design for the first time a new "1000 series machine" with China's independent, consistent command system and in line with American computer standards. In the end, with the joint efforts of many researchers, it took only one year and 2 months, and it looked like three DJS-130 computers with refrigerators side by side.
As China's first series of machines to form an industrial scale, it has produced more than 2,000 units in the future, and has been applied to many fields of national economy and national defense, which is an important milestone in the history of China's computer development.
While Chinese researchers are still familiar with a computer with just 13 boots assembled with various electronic components, it uses extremely traditional perforations to bring input and output results. The United States on the other side of the ocean has already stepped into the "PC era" in half a step.
Companies including HP and IBM have used miniaturized computer processors to reduce computers to the size of televisions, while introducing the concept of graphical interfaces on the basis of the original code system, from designing for private companies to computer concepts for personal use. Under the huge demand of personal computers, more and more new technologies and new products have come out, including Intel's first X86 processor 8086 launched in 1978, and Microsoft's first-generation Windows system MS-DOS launched in 1981.
Because of the huge difference in the investment in technology accumulation, China eventually "missed" the wave of the PC era. Since then, it has been the American and European companies that have led the development of the entire PC industry. Even if Lenovo eventually becomes the global PC sales champion with the "trade and technology" route, a considerable part of the profits still have to be taken by processor companies in Europe and the United States.
This situation has entered the "post-PC era" represented by smartphones and tablets to a certain extent. Because of the deep accumulation of the "(pre-)PC era", a number of companies, mainly in the United States and Europe, still dominated the early development of consumer electronics.
But this time we have three important chips: the first is the world's leading consumer electronics foundry capacity; the second is the world's largest and most popular mobile communication network; and the third is the world's most concentrated and willing to consume customers.
In the end, China's consumer electronics industry has made great strides in development driven by 3G, 4G and 5G. The most typical example is Huawei, in just over 20 years, from a small factory that produces telephone switches, it has finally become a giant in the global IT communication manufacturing industry, and at the same time, it is not weaker than or even more than the existence of companies in developed countries in many fields such as mobile phones, mobile phone SoC processors, 4G/5G technology, and communication base stations.
Although we finally caught up with the wave of technological development at the end of the "post-PC era", with the emergence of a new phase of the "era of artificial intelligence", new challenges have emerged – how china should ensure that it does not lag behind in the "era of artificial intelligence", or even lead the world.
The answer, in my opinion, is 3 words: root technology.
Disruptive artificial intelligence
Whether it is the "PC era" or the "post-PC era", the core of its lowest level is still the general computing power, that is, the CPU (central processing unit) is supported. In contrast, the era of artificial intelligence has undergone major changes in the entire technical system because of the overall change in algorithms.
This actually has to start from the principle, the core of the CPU is based on the semiconductor characteristics of the logic and calculation circuit, human programmers according to the CPU's binary algorithm characteristics, write out the CPU can efficiently logical judgment and calculation program.
Artificial intelligence, by contrast, is based on algorithmic "simulations" of humans, or rather simulations of how the human brain works. On this basis, combined with the logical judgment ideas of the human self for different tasks, a neural model is constructed, and then a large amount of real-world data is used to train the neural model, and finally a neural model can be applied and inferred.
Isn't it a bit abstract? We might as well compare two important milestones in general-purpose computing and artificial intelligence.
In 1992, IBM spent a lot of money to build the supercomputer "Deep Blue", using several cabinets and up to 480 special "chess processors" to defeat human chess masters. But in principle, what Deep Blue does is not complicated, it is still exhaustive, and it is not infinitely exhaustive.
At that time, human chess masters could probably calculate all the situations of the next 10 moves, while Deep Blue could calculate all the situations of the 12 moves. Deep Blue, who theoretically sees more future chess possibilities and corresponding advantages and disadvantages, and is less likely to make mistakes, will definitely win, but in fact loses to the human chess master in its first challenge. Later, after adjusting and optimizing for a year, I finally felt ashamed.
(Supercomputer "Deep Blue" first wins chess champion Kasparov)
In 2015, Google's AI startup DeepMind released the Go artificial intelligence AlphaGo, and directly competed for the then world Go champion. As the most complex board game of mankind to date, the theoretical move of Go is 10 to the 171st power. Each move has a variety of moves and will have an impact on the final outcome of the entire game. This is why humans have always believed that machines cannot defeat humans in Go by simply and exhaustively calculating.
However, the artificial intelligence that "armed" himself with deep learning ability was still unexpected, and he directly defeated the human world champion Li Sedol 4:1. More critically, in this game, the artificial intelligence showed a completely different strategy from the history of Human Go in the past 3,000 years. You can also say that artificial intelligence is far from just defeating humans, but in the continuous learning of Go, it has opened up a whole new set of moves with a higher winning rate.
Since the beginning of this competition, a new trend has begun to spread in the world's top Go circles, and these top players have adopted some of these thinking and moves after studying and learning the tricks of artificial intelligence. Take the Chinese Go player Ke Jie as an example, who once specifically said "thank You AlphaGo for the shock it has brought to our chess community".
In terms of core capabilities, universal computing may be "substitution + acceleration", and artificial intelligence may be "anthropomorphic + innovation". This is clearly not on one dimension.
What's more, after AlphaGo, more and more businesses are starting to apply AI to all walks of life. In translation, speech recognition, big data, autonomous driving, target recognition and other application scenarios have shown sufficient subversiveness, as long as there is a large enough data set, it can generate neural models and algorithms that far exceed the level of human programming, and finally achieve a computing effect far beyond human programming.
Obviously, artificial intelligence has always been a wave that China cannot miss.
Rooted in the ground, China is an important guarantee for ushering in the era of artificial intelligence
In fact, although China has made certain achievements in the development of the artificial intelligence industry, there are also hidden worries: compared with the "old players" with more abundant artificial intelligence foundations like the United States and Europe, China's accumulation of artificial intelligence root technology is much weaker.
According to the "2020 Chinese Intelligent Industry Investment and Financing Report" of the Head Leopard Research Institute, as of February 2019, the number of AI-related enterprises in China was 745, accounting for about 21.7% of the world, of which 67.3% were founded between 2010 and 2016. Most of the "young" Chinese AI companies were founded after the iconic AlphaGo GO AI incident in 2015.
AI companies as a whole are young, and the corresponding result is that such companies are more focused on the AI application layer, and very few companies are involved in the underlying AI root technology. Among the 745 AI-related enterprises counted above, 75.2% are application-layer enterprises, 22% are technology-level enterprises, and only 2.8% are located in the basic layer.
In terms of importance, the influence of "root technology" in the era of artificial intelligence will far exceed the PC era and the consumer electronics era. The so-called "root technology" refers to those technologies that can derive and support one or more technical clusters. Root technology is the root of the technology tree, which continuously provides nourishment for the entire technology tree, and largely determines the prosperity and death of the technology tree.
Although the whole process is also completed by computer hardware and software, because the entire computing logic is different from the traditional CPU and manual programming, the technology stack of artificial intelligence has many differences with the general computing represented by the PC.
On the whole, the technology stack of artificial intelligence is mainly divided into four parts, the lowest hardware infrastructure, the middle layer of software infrastructure, the upper layer of technology, and the top layer of application. Among them, the application and technology layers are more inclined to applications and solutions, which together as the application and technology layers. Among them, the hardware infrastructure part can also be divided into AI processors and AI hardware devices; software infrastructure can be divided into processor enablement, AI framework and development enabling platform.
The core existence of "root technology" lies in the two parts of "software infrastructure" and "hardware infrastructure". We can also see this from the layout of giants such as Nvidia and Google that entered artificial intelligence earlier.
Take Nvidia as an example, its GPU products were first developed by AI for deep learning training and reasoning, Nvidia in the continuous optimization of its own product AI operation efficiency at the same time, but also further deepened to the device layer, in addition to a variety of specifications and sizes of artificial intelligence GPU, but also specifically for different scenarios to create different solutions, there are business card size Jetson, there are also specifically for the automatic driving scenario of the Drive series of products, but also directly with the ultra-high-speed network to combine several GPUs into "super large" GPU DGX.
In terms of software infrastructure, Nvidia's CUDA solution is even more far-reaching, and in the AI framework, NVIDIA directly adopts Google's TensorFlow and Facebook's PyTorch. This is because NVIDIA has chosen to focus more on the technical layer and deepen the application of the SDK into the industry.
This is followed by Google, which not only encourages subsidiary DeepMind to advance the AlphaGo project. At the same time, in order to provide AlphaGo with sufficient computing power, Google has also developed a TPU processor dedicated to artificial intelligence. And the TPU host server, into its own cloud service system. In the following years, the TPU processor version and its solutions were continuously updated, and finally TPU was opened to a wide range of customers as a cloud service business content.
In addition to the efforts of the two companies in the "root technology", there is also a point that is particularly noteworthy, that is, the technical layout collaboration of the artificial intelligence full-stack road: NVIDIA uses CUDA to string together its best GPU hardware and the entire software architecture and ecology on top of it, and Google has made the industry's most popular AI framework TensorFlow according to the rich accumulation of its own artificial intelligence technology.
Looking at Google and NVIDIA, the two leading companies in the global artificial intelligence industry, they have invariably chosen to lay out the key nodes of AI root technology at the same time, and through the collaboration between key nodes, they can maximize their artificial intelligence ecological capabilities and efficiency.
Most importantly, the artificial intelligence ecology of core companies will also radiate to the entire country and even the global artificial intelligence industry over time, forming a potential voice for companies and countries in the artificial intelligence industry.
China's development of artificial intelligence industry still needs to "overtake in curves" or even "strike late", and the only key to the next is to develop its own artificial intelligence.
Breaking the game of artificial intelligence, what should Chinese companies do?
The most important and the biggest shortcoming of Chinese enterprises at present is the basic hardware of artificial intelligence. More specifically, it includes AI processors and various solutions built with processors.
There are three main reasons, one is that AI hardware is the main boundary of AI application promotion, especially scenarios such as smart phones, Internet of Things, smart cities and other scenarios that emphasize end-side data acquisition and processing, which often require tailor-made AI computing capabilities and compact solutions; secondly, the processor architecture and development methods must achieve a high degree of uniformity; and finally, the independent innovation of basic hardware.
In addition, in order to maximize the value of these basic hardware, it is also necessary to support it with efficient software infrastructure, mainly including "chip enablement", "AI framework", "development enabling platform".
It is obviously not easy to make progress or even breakthroughs in so many links in one go, but many Chinese enterprises have still made a lot of achievements in these years. A large number of AI processors and AI algorithm companies have come into being, and have begun to fill the gap in AI root technology, for example, many new car-making forces are now replacing foreign AI processors with domestic startup products.
There are more progress in the AI framework and development end, not only the long-term heavy layout of Internet giants such as Baidu Tencent, but also the subdivided application scenarios, such as iFLYTEK in speech semantics, and the "AI Four Tigers" in the field of smart city applications.
Starting from the pursuit of layout integrity and forward-looking, the biggest progress is made by Huawei. Huawei's artificial intelligence is a development path with "root technology" as the core grasp. At present, ascend computing industry from the basic hardware to the basic software layer has formed a full-stack and all-scenario "root technology" layout, and in the cloud, edge, end side are deployed a unified architecture of ascend series solutions, its basic software layer of heterogeneous computing architecture CANN and AI development framework MindSpore can not only optimize the process, but also with the ascend hardware foundation for in-depth optimization and integration, full-stack tuning. On top of these root technologies, the AI development platform MindX has also further accelerated the development and deployment of AI applications.
In China's AI era, there can be no "roots". With the deepening of the construction of the Chinese intelligent industry, we will also see more and more Chinese AI companies deeply engaged in the products and businesses of artificial intelligence "root technology". The development of independent innovation "root technology" must be the consensus of China's construction of artificial intelligence industry. Although we will still face challenges in the development of breakthrough artificial intelligence, Chinese AI companies have proved with the efforts of the past few decades that in the global AI root technology competition, China also has the opportunity to catch up and even surpass.