With the rapid development of the AI emerging industry, traditional chips can no longer meet the requirements of the AI industry on-chip performance and computing power. Therefore, building efficient AI chips and effectively combining chip technology with AI technology has become a hot topic. EqualOcean believes that AI chips, as the foundation and core of AI and related applications, are fields with potential. In the first quarter of 2022, EqualOcean released the 2022 China AI chips Industry Research Report, which makes a detailed analysis of the mainstream types of AI chips, shows the development status of China's AI chips industry, and explores its development difficulties and opportunities.
Trends in the AI chip industry
AI algorithm needs to be implemented on computer equipment, and the chip is the core part of computer equipment operation. The development of AI chips mainly depends on two fields: the first is the mathematical model and algorithm established by imitating the human brain, and the second is the semiconductor integrated circuit, namely the chip. Advanced algorithms need enough computing power, that is, the support of high-performance chips. The development of AI on-chip is divided into three stages: in the first stage, due to the insufficient computing power of the chip, the neural networks algorithm failed to launch; In the second stage, the computing power of the chip is improved, but it still cannot meet the requirements of neural networks algorithm; In the third stage, GPU and AI chips of the new architecture promote the launching of AI. With the emergence of the third generation of neural networks, China has gradually filled the barrier between neuroscience and machine learning, and AI chips are developing closer to the human brain.
In 2021, the 14th five-year plan (2021-2025) for national economic and social development and the long-range objectives through the year 2035 pointed out that during the 14th five-year plan (2021-2025), China's new generation AI industry will focus on key fields such as high-end chips. It has established a good policy environment for the AI chips industry from a national perspective. According to their backgrounds, all localities have also issued plans to promote the AI chip industry. By September 2021, more than 20 provinces, including Beijing, Tianjin, Shanghai, Jiangsu, and Fujian, had issued AI-related policies to support further and guide the AI and chip industry development.
In terms of industry data, compared with 2020, the number of investments in the AI field has decreased, but the scale of single investment shows an upward trend. The AI chips industry also continues to have capital entry, and the amount of single financing exceeds CNY 100 million. As of January 2022, there were 92 financing events in AI chips related fields in China in 2021, with about CNY 30 billion. Policy support and market demand are still the main driving forces for developing AI chips. According to the calculation of EqualOcean, in 2025, the market scale of China's AI core industry will reach CNY400 billion, of which the market scale of fundamental layer chips and related technologies will be about CNY 174 billion.
Decomposing the AI chip industry
1. Technical layer
AI chips can be divided into GPU, FPGA, ASIC and brain-inspired chips according to their technical architecture; According to their location in the network, they can be divided into cloud AI chips, edge and terminal AI chips; According to their goal in practice, it can be divided into training chip and information chip.
AI hardware acceleration technology has gradually matured. In the future, more innovations may come from the combination of circuit and device level technologies, such as in-memory processing and brain-inspired computing, or for unique computing modes or new models, such as spark matrix computing and approximate computing; Or optimize the architecture for the characteristics of the data rather than the model. At the same time, if the algorithm does not change significantly, according to the main methods of AI acceleration and the development trend of semiconductor technology, it will reach the limit of a digital circuit soon. It will rely on approximate, analogue, and even materials or fundamental research innovation.
2. Application layer
With the maturity of technology, the application scenarios of AI chips will be in the cloud and big data centre and move to the edge with the computing power and be deployed in smart homes, smart manufacturing, digital finance, and other fields. At the same time, with the increasingly rich types of smart products, smart terminals, smartphones, security cameras, and self-driving cars will become more and more popular.
At present, most AI training and reasoning workloads occur in the public cloud and private cloud, and the cloud is still the centre of AI. Driven by the demand for privacy, network security and low latency, AI training and reasoning workload on gateways, devices and sensors appear in the cloud. Higher performance computing chips and new AI learning architecture will be the key to solving these problems. The Internet is an industry with strong demand for computing power in the cloud. Therefore, in addition to traditional chip enterprises, chip design enterprises and other participants, Internet companies have joined the AI chips industry to invest in self-developed cloud AI chips.
According to the calculation of EqualOcean data, the scale growth of China's automatic driving industry will reach 24% in 2022, and the shipment growth of intelligent camera products will exceed 15%. Shipments of smart products such as mobile phones, tablets, and VR / AR glasses also increased significantly, increasing the demand for smart chips. At the same time, the types of intelligent terminal products are also gradually diversified. Consumer hardware such as intelligent audio, service / commercial robots, industrial/numerical control equipment and communication products is becoming more abundant. Different product types also put forward more requirements for chip performance and cost.
Opportunities and Challenges
In chip design and manufacturing, China still lacks design software, and there is still a gap between advanced processes and equipment and the world's top level. Some products and equipment in this field are still very dependent on imports. The quantity and quality of data determine the accuracy of the AI model. At present, most of the data generally belong to different institutions or departments, such as government departments, the financial industry or the medical industry. It is challenging to integrate them into a whole, which has caused significant obstacles to improving AI technology.
After fully recognizing the importance of data, local governments set up big data management bureaus to effectively use data insecurity, government affairs, legal affairs and other fields from the government level. At the same time, formulate better data management policies to make data better serve the local economy and effectively solve the problem of data island. Since the outbreak, people have paid more attention to the network and accumulated more data. In 2021, the big data of leading Internet companies reached thousands of Pb level, the data volume of leading enterprises in traditional industries also reached Pb level, and the data generated by individuals reached TB level. Currently, China accounted for 23% of the global data volume in 2018 and is expected to reach 27.8% in 2025.
The transformation direction of China's digitization has driven the gradual improvement of the underlying technology, and the international influence is also rising year by year. At the same time, it has gradually established a dominant position in big data, chip design and application launching. Industrial development has also attracted more overseas talents to return home for entrepreneurship and employment. The industrial chain structure may be reconstructed in the future, and more enterprises, universities, and organizations may form a joint force to promote the new development of AI and chips jointly. At the same time, AI-related applications will also be widely used to develop health care, education, media, finance and customer service. Manufacturing and transportation will be significant application scenarios of AI.
Bottom line
Overall, the development of AI chips still needs the accumulation and precipitation of basic science. Therefore, integrating industry, University, and research is an effective way. Make full use of enterprises, universities, scientific research institutions and other different educational environments and educational resources, combine the teaching of theoretical knowledge with industrial engineering practice and scientific research practice, cultivate and accumulate high-quality talents in the field of AI, and maintain the sustainable development of China's AI and chip industry.
Check the full report here.