[ Security Exhibition Network Market Analysis ] The AI chip industry is just in its infancy and the market is growing rapidly. The traditional chip industry is already a mature industry. Traditional chip design and wafer manufacturing packaging and testing are both serious technical barriers and the market is growing slowly. The artificial intelligence industry is in its infancy, and some AI products can already land and continue to be optimized. However, the algorithm gradually stabilizes.
First, the definition and main types of AI chips
At present, there are many AI chip design schemes. In a broad sense, all chips for artificial intelligence applications can be called AI chips. Neural network algorithms represented by "deep learning" for AI operations. The system needs to be able to efficiently process large amounts of unstructured data (text, video, images, speech, etc.), so the hardware needs efficient computing capabilities. The AI chips that perform artificial intelligence tasks can be divided into four architectures. According to the perspective of artificial intelligence system development, they can be further divided into two categories. The first category is CPU and GPU, which are called software and hardware. The second category is FPGA and ASIC are called hardware and software.
Figure: Comparison of four main types of AI chip architecture, data source: OFweek Industrial Research Institute
According to the application scenario, it can be divided into server-side (cloud) chip and terminal (edge-side) chip. During the training phase of deep learning, due to the extremely large amount of data and calculations, a single processor cannot independently complete the training process of a model. Therefore, the chip responsible for the AI algorithm uses high-performance computing technology. It must support as many networks as possible Structure to ensure the accuracy and generalization of the algorithm. On the other hand, it must also support floating-point operations. At the same time, in order to improve performance, it must support array-type structures and perform accelerated operations. In the inference stage, since the trained deep neural network model is still very complicated, the inference process is still computation-intensive and storage-intensive, and can be deployed on the server side.
The edge-end and server-side AI chips are essentially different in design thinking. The edge-end AI chips must ensure high computing energy efficiency. In application areas where advanced assisted driving ADAS and other equipment have high real-time requirements, the inference process must It is completed in the device itself, so the mobile device is required to have sufficient inference capabilities, and some applications require the edge chip to have low power consumption, low latency, low cost and other requirements, resulting in a variety of edge AI chip products. Style.
Second, the global artificial intelligence chip market size
The development of artificial intelligence and blockchain has led to the rapid growth of special application chips. The development path of artificial intelligence chips has gone from general to special. According to the annual report of the development of China's Internet of Things
, the global artificial intelligence chip market size reached 2.4 billion in 2016. The US dollar is expected to reach 14.6 billion U.S. dollars by 2020, with rapid growth and huge development space.
Figure: Global artificial intelligence chip market size and forecast (USD 100 million) from 2016 to 2020, data source: China Internet of Things Annual Report, OFweek Industrial Research Institute
In terms of market forecasts for segmented fields, you can observe from the research data of CICC. In 2017, the overall AI chip market size reached US $ 6.27 billion, of which US $ 2.02 billion was used for cloud training AI chips, US $ 340 million was used for cloud inference chips, and AI chips for edge computing were 39.1 By 2022, the overall AI chip market size will reach 59.62 billion U.S. dollars, CAGR57%, of which cloud training AI chips are 17.17 billion U.S. dollars, CAGR53.5%, cloud inferred chips are 7.19 billion U.S. dollars, CAGR 84.1%, edge computing AI The chip is 35.22 billion US dollars, CAGR is 55.2%.
Figure: 2017-2022 AI chip segment market forecast (USD 100 million), data source: China Internet of Things Annual Report, OFweek Industrial Research Institute
Industry chain and business model of the AI chip industry
In the AI chip industry's industrial chain and business model, although AI chips are emerging fields, they are in the market for strong rivals. The integration of upstream and downstream industry chains has already begun. The semiconductor industry chain is long and has the characteristics of both high capital and technological barriers. The upstream is mainly chip design. According to the business model, it can be further subdivided into IP design, chip design foundry and chip design. The midstream includes wafer manufacturing and package testing. The downstream of the industrial chain is divided into sales and system integration companies. Artificial intelligence solution providers that provide software and hardware integration solutions are classified as system integrators.
The overall semiconductor business model is divided into two types. The first is the vertical integration model (IDM). The business of this model includes design and manufacturing / package testing. The representative companies of the IDM model are Intel and Samsung. The second is the vertical division of labor model. Enterprises adopting the division of labor model only specialize in one business. For example, Nvidia and Huawei Hisilicon only have chip designs, also known as fabless, while TSMC and SMIC have only foundry manufacturing operations, called foundry.
There are three business models for chip design: IP design, chip design foundry and chip design. Most of the artificial intelligence startups are mainly in the field of chip design, but there are many traditional chip industry leaders in this field, such as Nvidia, Intel, Xilinx, and NXP, so only a few AI chip designers will enter the field of traditional chip companies to compete with, including Cambrian and Nvidia competing in the server chip market, and Horizon competing with Nvidia and NXP for autonomous driving chips market.
Figure: AI chip industry chain, data source: OFweek Industrial Research Institute
4. China's AI chip industry policy environment
In July 2015, the State Council proposed guidance on upgrading the horizontal connectivity of industries with “Internet +” as the core. Then in the “Robot Industry Development Plan” released in April 2016, artificial intelligence has gradually become the core project for policy development. The "New Generation Artificial Intelligence Development Plan" put forward in March set the strategic goals for the three phases of 2020, 2025 and 2030. The first phase of the "Three-year Action Plan to Promote the Development of the New Generation of Artificial Intelligence Industry (2018-2020)" ", Will focus on supporting neural network chips, and look forward to mass production and large-scale application of AI chips.
Figure: Policy environment of the AI chip industry, data source: OFweek Industrial Research Institute
V. Competition pattern of global AI chip manufacturers
Overview of AI chip manufacturers at home and abroad. Participants in the AI chip field include traditional chip design, IT manufacturers, technology companies, networks and start-ups. The products cover CPU, GPU, FPGA, ASIC, etc. In the AI Chipset Index TOP24 list released by Compass Intelligence in 2018, the top ten companies are still European, American, Korean, and Japanese companies. Domestic chip companies such as Huawei Hisilicon, MediaTek, Imagination (acquired by China Capital in 2017), Cambria Ji, Horizon and other companies all entered the list, of which Huawei Hisilicon ranked 12th, Cambrian ranked 23rd, and Horizon Robotics ranked 24th.
Chip design companies are still the main force in the current AI chip market. Major manufacturers include Nvidia, Intel, AMD, Qualcomm, Samsung, NXP, Broadcom, Huawei Hisilicon, MediaTek, Marvell, Xilinx, etc., and also do not directly participate in the chip Design, ARM chip company only authorized by the chip IP.
6. Financing status of major new AI chip manufacturers in China
The following table lists the financing status of China ’s major new AI chip makers. Both Horizon B and Cambrian are Unicorn companies with valuations of US $ 2.5 billion and US $ 3 billion. In the case of corporate mergers and acquisitions, in July 2018, Xilinx, the global leader in FPGAs, acquired the domestic AI chip design manufacturer Shenjian Technology, marking the official start of the battle for resources to increase the valuation of the AI chip industry. The two parties have joined forces not only to provide cutting-edge machine learning solutions for global customers, but also does not have to bear high chip design and R & D costs alone. Chips from development to finished product IP licensing, development software, tape-out, chip manufacturing / package testing Such costs are unavoidable development costs, and Xilinx has also won chips that compete with giants such as Nvidia, Intel, and Google in the field of AI chips to achieve a win-win situation.
Figure: Products and layout of global AI chip manufacturers, data source: OFweek Industrial Research Institute
Figure: Financing status of major new AI chip manufacturers in China, data sources: corporate public data, OFweek Industrial Research Institute
Development Trend of China's AI Chip Industry
The main landing markets of AI chip companies include cloud (including edge) servers, smart phone mobile terminals, IoT terminal devices, and autonomous driving. These markets are all tens of millions of shipments or tens of billions of dollars in sales. The market is huge. At present, a number of overseas and domestic technology giants and startup companies have made certain achievements in product development and marketing, such as traditional chip established companies such as Nvidia, AMD, IBM, Intel and Qualcomm, as well as Apple, Google, Amazon, etc. Internet companies such as Huawei, Alibaba, and Baidu have gradually emerged as newcomers such as the China Cambrian, Horizon, etc., and their products have gained prominent applications in cloud, autonomous driving, smart security, and mobile Internet scenarios.
The AI chip industry is just in its infancy and the market is growing rapidly. The cooperation model between chip companies and customers is still being explored. At the same time, there is also a bubble in the AI chip industry. The market looks forward to the development of companies with technical strength and performance. Under the background of the development of high-end semiconductors in the country, a large number of AI chip industry companies have emerged in the past two years, proving capital and industry employees. Recognizing the future application prospects of AI chips, the other side also shows that the technical threshold of AI chips is not as high as that of CPUs, or that the technical threshold of low-end AI chip products is not high. However, it should be noted that the high cost of chip research and development and manufacturing and the great demand for funds are also the characteristics of this industry. It is expected that in the next 2 years, with the first batch of AI chip products from various manufacturers coming on the market, the market will test the products and technologies of various manufacturers. The teams with insufficient technology and uncompetitive products will continue to lack support from subsequent orders and profit. Withdrawing from the market, the only companies that can survive are the market-recognized and technically strong teams.