Intel’s home game is PC, Qualcomm’s home game is mobile internet, whose home game will AI be? In 2016, AlphaGo’s defeat of Li Shishi was a battle for AI’s fame. Since then, it has been out of control. AI has become a “just-needed” for all terminals overnight, and products only have a selling point when they are loaded with AI functions. Subsequently, terminal equipment manufacturers quickly opened the AI ​​entrance to grab the battle, after some competition, the smart speaker achieved a phased victory. According to data from Strategy Analytics, in 2017, the global shipment of smart speakers reached 32 million units, and the shipment volume in 2018 increased to 86.2 million units.

However, as terminal devices evolve to AI, AI requirements are severely fragmented, and different devices have different computing power requirements for AI, and it is difficult for the original general-purpose architecture chips to cover all of them. With the access of AI algorithms, especially for deep learning, whether it is CPU, GPU, FPGA, or heterogeneous architecture, they cannot meet the parallel computing capabilities and higher storage bandwidth required by terminal devices. AI dedicated chips are increasing in computing density and There is an absolute advantage in power consumption.

Domestic manufacturers are rushing into the game, self-developed AI chips are enthusiastic

2017 is the year of rapid advancement of AI technology and the year of the outbreak of AI chips. Many domestic manufacturers have entered the AI ​​chip market. Corresponding AI chips have appeared in various applications of terminal equipment. “Let professional products do “Professional things” have become the consensus in the industry, and only AI chips can better realize AI functions.

Among them, Yunzhisheng launched the AI ​​chip for IoT interaction scenarios-Swift, Bitmain launched Sophon Sophon BM1680 for tensor computing acceleration, and Horizon launched Journey 1.0 and Rising Sun 1.0 for smart driving and smart cameras, respectively. . Huawei released the Kirin 970, which is the first to incorporate a neuron network unit (NPU) to complete artificial intelligence calculations.

From fever to fever, how has the domestic AI chip market gone through?

Although there are not many companies that release AI chips, the number of domestic IC design companies has grown to more than 1,600 in 2017. Many companies are catching up with the AI ​​craze and are preparing to enter the AI ​​chip market.

The feast period of the domestic AI chip market

Turning to 2018, the AI ​​craze is booming, and it coincides with the ZTE incident. While domestic chips are touted in the hot search, the development of the AI ​​chip market is also at a climax, and traditional chip giants and startups are beginning to actively expand their layout. . Baidu released the AI ​​chip-Kunlun, which claims to be based on Baidu’s CPU, GPU and FPGA accelerator. After 8 years of research and development, it was launched after more than 20 iterations. It is positioned as a cloud-based full-function AI chip; Huawei released an upgrade. Both Teng 310 and Shengteng 910 use Huawei’s self-developed Da Vinci AI architecture, covering AI inference and AI training. They are positioned as the world’s first AI IP and chip series covering all scenarios, with the ability to cross the cloud and edge , The best energy efficiency ratio of the whole scene, and proposed to make full-stack and all-scenario AI solutions, and decided to “take everything” in the AI ​​industry chain.

From fever to fever, how has the domestic AI chip market gone through?

For terminal manufacturers, it seems that only the release of an AI chip can prove the company’s technical strength and its own product characteristics. Ask and Rokid, who have been developing terminal products, also joined this AI chip war. In 2018, Qiaowen released two AI chips—Wenxin Mobvoi A1 and Wenxin Mobvoi B1. Rokid released KAMINO18, and both companies developed products in cooperation with Hangzhou Guoxin Technology. Hangzhou Guoxin Technology released AI chips GX8010&GX8008 in October 2017, with digital signal processor DSP, neural network processor NPU, and standard interfaces such as USB/IIS/IIC/UART. When asked when going out, Wenxin has made in-depth optimizations for artificial intelligence and the Internet of Things, so that various IoT devices have low power consumption and strong offline AI capabilities.

Rokid founder and CEO Misa clearly stated, “The current chip is basically SoC, and 90% of the things in SoC are very mature. Rokid does not need to spend energy on various types of IP. How to integrate Rokid’s algorithm and how to optimize at the SoC architecture level. Rokid does not make money from chips, we do not directly sell chips separately, and Rokid does not make chips based on the starting point of making chips, just because there is no chip on the market. We need what we need, so let’s do it.” In addition, because Rokid has its own terminal equipment, KAMINO18 has already received orders for one million pieces when it was released. The ability to design AI chips has obviously become a highlight of Rokid.

The boom in AI chips has also stimulated the activity of the capital market. In 2018, many AI chip companies received financing, Cambrian received hundreds of millions of dollars in Series B financing, and Yunzhisheng received 600 million RMB in C+ round financing. Chi announced that it has received 500 million yuan in financing and announced the promotion of the implementation of AI chips, etc., and Horizon announced that it will complete a new round of 500 million to 1 billion US dollars in financing (actually, 600 million US dollars in financing will be completed in 2019).

AI is starting to cool down, what should AI chip manufacturers do?

There will be tides and tides. After the carnival in 2018, investors have become more calm about AI investment in 2019. Coupled with the lack of optimism in the international trade environment and the semiconductor market, AI chip entrepreneurs are facing huge challenges. Investors pay more attention to the landing ability of startup companies. They divide AI chip companies into two categories, one is companies with a certain scale, and the other is start-up companies. They believe that the former can introduce AI chips into the existing system in accordance with the needs of customers, and the products will be easier to land; however, startups lack landing applications, and the cost of chip design is high, and the mass production cycle of chips is relatively long. It can only be maintained by financing. Once the capital chain breaks, it will be unsustainable.

It can be seen from the release of new products that in the first nine months of 2019, only Ali Pingtou released a high-profile Hanguang 800, which is mainly used for AI reasoning; Bitmain’s Suanfeng BM1682 did not carry out publicity; Ziguang Zhanrui’s Tiger Ben T710 and Huben T618 are not AI dedicated chips, but only have AI functions. The AI ​​chip market has entered a cold winter.


In 2017, the blowout growth of smart speakers attracted the attention of many chip companies. The explosion of voice application scenarios led to the accelerated rise of AI-specific chips, and the voice market was used as a node to spread to more application markets; in 2018, domestic manufacturers AI The enthusiasm for core-making has reached a climax. 1,600 domestic companies have begun to pay attention to the AI ​​chip market, but the applications that actually landed are mainly concentrated in smart phones, intelligent voice control and video analysis; in 2019, AI began to cool down, and the AI ​​chip market began to squeeze bubbles. Only companies with sufficient capital and technology are more likely to go on.

With the continuous segmentation of AI application scenarios, terminal manufacturers will inevitably prefer to choose corresponding AI dedicated chips and highlight their own product features. The application prospects of AI chips are still promising. The future market trends can be summarized as follows:

First, the implementation of AI algorithms needs to rely on chips, and different algorithms have different requirements for chips. For specific algorithms, the acceleration of dedicated AI chips is much better than that of general-purpose chips, so more and more terminals Manufacturers began to develop AI chips to better meet their own product needs;

Second, AI dedicated chips are more targeted. From the products released in the past three years, we can see that Horizon’s products are aimed at autonomous driving and visual analysis, and Rokid is aimed at intelligent voice control, Baidu and Ali are aimed at cloud computing, and more dedicated AI chips will appear in the future;

Third, in terms of market development, it is relatively easy to extend from terminal products to AI chip research and development, which can not only ensure the landing of applications, but also highlight product features. For example, the AI ​​chips developed by Rokid and Rokid for intelligent voice products make up for the shortcomings of general-purpose chips and highlight the advantages of their own voice products;

Fourth, Internet giants will inevitably embark on the journey of AI core making. The release of AI chips based on their own architecture by Baidu and Ali is the best proof. After all, the products developed by general companies cannot cover the application needs of these giant companies.

The Links:   FF800R17KE3-B2 PHT2508CF