After years of development, artificial intelligence has emerged in the revolution of new science and technology and industrial innovation, and has become a frontier strategic area in which developed countries and regions stimulate innovation dynamism, seize the opportunities for future development, and consolidate core competitiveness. Basic breakthroughs in artificial intelligence basic research, brain-like research, human-computer collaborative enhancement intelligence, and community integration intelligence have represented major breakthroughs in the future of science and technology and industrial development, reflecting the current development trend of smart economy and green economy in the world.
The rapid development of the artificial intelligence industry has benefited from the massive increase in data and the continuous improvement of computing power under the influence of Moore's Law. Whether it is the acquisition or storage of massive data or the ability to calculate, it is inseparable from the hardware carrier. . Therefore, artificial intelligence chips have become a very strategic link in the current fierce artificial intelligence industry competition, and are also one of the areas most concerned about many funds invested in artificial intelligence in the past two years. At the same time, this is also a huge market that is growing rapidly. Morningstar, the international authoritative fund rating agency, predicts that the global AI chip market may exceed 20 billion U.S. dollars in 2021.
There is no accurate definition of artificial intelligence chips so far. In a broad sense, chips that meet the needs of artificial intelligence applications can be called artificial intelligence chips. In fact, at present, most of the artificial intelligence application scenarios, we still use GPU, FPGA and other existing general-purpose chips for parallel computing to implement artificial intelligence algorithms. However, reviewing the history of computing chip development, new computing models will generally generate new dedicated computing chips, so in the future, artificial intelligence dedicated chips will be the general direction of development, and will produce disruptive changes to the traditional computing architecture, which is also Why the artificial intelligence chip has a strategic reason, its development is not only significant for the artificial intelligence industry itself, but also has a profound impact on the current global chip industry's market structure. For China, it may be a major catch-up. opportunity.
At present, the two countries with the most fierce competition in global artificial intelligence chips are the United States and China. Of course, they are not just chips. The two countries are all-round competing in the field of artificial intelligence. Today, they say only that they are chips.
The competition for artificial intelligence chips is currently divided into three major tracks:
The first is a semi-custom solution based on general-purpose chips such as GPUs and FPGAs. For example, Nvidia's development of GPUs for various types of smart computing devices, and the creation of the NVIDIA CUDA platform have greatly increased its programming efficiency, openness, and richness, and established algorithms including CNN, DNN, depth-aware networks, RNN, LSTM, and reinforcement learning networks. The platform allows AI to penetrate into various types of smart machines. The U.S. has inherited the strong advantages of traditional computing chips, including the world's first technological capabilities and industry status in the areas of CPU, FPGA, GPU, and DSP, naturally occupying the dominance of this track, allegedly in medical and life sciences. There are nearly 4,000 companies that use GPUs or FPGAs to conduct in-depth learning in energy, financial services, automotive, manufacturing, and entertainment industries. The penetration of traditional advantages is evident. Can also be seen from the side, in this track, in fact, there are few opportunities for Chinese enterprises.
The second track is a dedicated chip for deep learning algorithms. Although general-purpose chips such as GPUs and FPGAs are suitable for large-scale parallel computing, they also have bottlenecks in performance, power consumption, etc. In the face of continuously increasing data volume and continuously expanding AI application scale, general-purpose chips naturally have their limitations. Therefore, dedicated chips must be the trend of the times. Dr. Chen Yunzhao, a researcher at the Institute of Computing, the Chinese Academy of Sciences and the founder of the Cambrian Deep Learning Processor chip, wrote in the "Communications of the Chinese Computer Society": It is possible to design specialized instruction sets, microstructures, artificial neuron circuits, and memory hierarchies. Increase the efficiency of intelligent processing of deep learning models by 10,000 times in 3 to 5 years. This also makes most of the companies in the China-US artificial intelligence chip startup focus on this track to compete. At present, this track gathers the most active intellectual resources in the world and brings forth a variety of methods to customize the chip design and architecture to solve the “pain points†of artificial intelligence chips in different AI application scenarios, such as affecting processor performance improvement. The memory bandwidth bottleneck, the performance of computing unstructured information is weak, and for example, in the implementation of AI applications on embedded devices, in addition to the requirements of computing performance, how to balance power consumption and cost issues. In this track, there are huge investments in global artificial intelligence and chip giants such as Google, Intel, and Nvidia, and numerous small and medium-sized start-ups are working hard to expect the traditional giants to overcome the sturdy computing architecture. Barriers to create Intel or ARM in the age of artificial intelligence. There is a large number of AI start-ups from China. In the investment in AI, China has surpassed the United States as the No. 1 in the world, and a large amount of funds have been invested in the field of AI-specific chips, which fully proves this track. There is an extraordinary significance for China's computing chips.
The third track is a brain-like computational chip. This area is no longer confined to the acceleration of deep learning algorithms, nor is it confined to the realization of artificial intelligence in specific scenarios, but it is hoped that the basic structure of the chip and even the device level can be developed. The new non-von Neumann computer model and architecture to solve the problem of general-purpose intelligent computing. There is still a gap between the brain-like research and the mature commercial technologies that can be widely used on a large scale, and even in the process of industrialization, there is still a considerable risk, but in the long run this track is most likely to bring about a revolution in computing systems. s Choice. In fact, this area is more of a competition for basic research capabilities. The gap between China and the United States is actually not great.
Artificial intelligence chips will determine the infrastructure and future ecology of the new computing era. As a result, U.S. giants such as Google, Microsoft, IBM, and Facebook have invested heavily in accelerating the research and development of artificial intelligence chips to capture the strategic commanding heights of the new computing era. Artificial intelligence era dominates. The domestic artificial intelligence chip presents a pattern of innovation and activity and flourishing flowers. Through visits to research, communication with a number of investment institutions, and startups, here are currently close to 40 domestic companies involved in artificial intelligence chip business as a sample, summing up some of the characteristics of the current development of artificial intelligence chip industry in China:
First of all, Beijing is the most active area of ​​innovation in artificial intelligence chips in China, and the number of companies involved in this business is more than half. Beijing has the most intensive intelligence resources in the field of artificial intelligence and microelectronics in China. Tsinghua University, Microsoft Research Asia, the Chinese Academy of Sciences Institute of Computation, Institute of Automation, Microelectronics Institute, Beihang, and more than half of the country’s artificial intelligence research units are all clustered in Beijing. In addition, Baidu, Xiaomi, Jingdong, Didi and other domestic Internet giants have invested fully in the field of artificial intelligence, as well as active investment institutions and industry media boosting catalysis, making Beijing the most fertile ground for the growth of artificial intelligence chip startups. According to the 2017 “China Artificial Intelligence Industry Development City Ranking†released by the Yiou Think Tank, Beijing’s scores on enterprise scale, policy basis, academic foundation, etc. score far more than other cities, ranking first, and also showing that Compared with other cities, Beijing has a unique advantage in developing the artificial intelligence chip industry.
Secondly, the current domestic scene of artificial intelligence chip companies is mainly in the field of security applications. Among the 38 domestic artificial intelligence chip companies in the statistics, nearly 30 companies with security as the core business are facing auto-driving, medical, and smart home automation. Other scenes are relatively few in the development of AI chips. The security field has a huge amount of data, which can provide enough scenes for deep learning and training. At the same time, this field can meet the needs of AI chips for large-scale and high-capital investment, that is, the quantity, but also the money, and the security market. The fragmentation characteristics and geographical closedness properties are presented. It is relatively good for AI chip companies to enter business operations, and is naturally the preferred market for them.
Third, we have found that many old chip companies are actively embracing changes in the era of artificial intelligence, approaching 1/4, becoming an important force in the Chinese artificial intelligence chip industry. Hais, Zhongxingwei, Beijing Junzheng, Zhongtianwei, Hangzhou Guoxin, and other traditional SoC processor chips or multimedia chip companies established before 2013 are representatives. Compared to those active start-up AI chip newcomers, veteran companies have more complete front-end design, product, verification and test teams, and have the engineering experience to build a complete SoC chip product, knowing that only one deep learning accelerator is Can't use it. The entry competition of veteran chip companies and the wrestle of the freshman force at the talent and product level are the highlights of the future industry.
Fourth, in the frontier innovation field of some AI chips, the technical capabilities of some of our companies have reached the international advanced level. On the one hand, it is embodied in basic research, such as the field of in-memory computing where computations are performed directly in memory without data transfer, heterogeneous fusion brain computing, reconfigurable computing chip technology, and basic research capabilities in China Connected or even led. Here, we must mention that Beijing's future chip technology advanced sophisticated innovation center in Tsinghua University has done a lot of work in the frontier areas of artificial intelligence chips, with a certain degree of global influence. On the other hand, it is reflected in the large number of overseas AI chip experts and high-end talents returning to China to start their own businesses. According to LinkedIn data, from 2013 to 2016, the average annual growth rate of AI graduates returning from home is about 14%, while there is overseas work. The average annual growth rate of returned AI personnel in the background is about 10%. Among the 38 AI chip companies we count, there are more than half of the teams with overseas background.
Fifth, there has been an increase in the number of algorithms and system companies that have entered the field of artificial intelligence chips, such as Shangtang, Haikang, Yitu, and Huawei. Algorithm vendors provide high-frequency, basic functional services. Therefore, achieving revenue through algorithms alone is often a bottleneck. By chipping their respective artificial intelligence core algorithms, it will not only improve the original performance, but also paves the way for commercial profits. The system companies are also actively investing in AI chips through independent R&D and M&A investment. They are far from the scene, they have a deep understanding of the real needs of the scene, they have powerful hardware and software integrated product capabilities, and they have marketing channel capabilities. Adequate capital reserves are the advantages of system enterprises, which makes them always in a more dominant position in dealing with the relationship of many start-up companies with AI chips, or work together, or compete directly, increasing the uncertainty of the industrial structure.
In short, at present, the global artificial intelligence industry is still undergoing rapid changes. We have plenty of reasons to be optimistic about China's potential in the field of artificial intelligence. The broad industry distribution provides a broad market for the application of artificial intelligence, but compared with the market and data. In terms of advantages, the development of artificial intelligence chips in China is still struggling to catch up. Although there are breakthroughs on different technical routes, there is still a long way to go.
Notonthehighstree Notonthehighstree
Bossgoo(China)Tecgnology.(Bossgoo(China)Tecgnology) , https://www.cn-gangdao.com