The chips of Chinese medical AI companies are all from the United States. Will they be "stuck in the neck" like ZTE?
In this chip battle, will the war spread to the artificial intelligence industry, especially the medical artificial intelligence industry that we are concerned about. If Sino-US trade frictions escalate again, will medical artificial intelligence companies be "stuck in the neck" like ZTE?
The booming "ZTE Incident" is continuing to ferment. The US Department of Commerce issued a ban on export authority to ZTE is not only a chain reaction caused by the escalation of the Sino-US trade war, but also highlights the high-tech industries of China and the United States The power gap.
In modern warfare, science and technology games will be an important force. China and the United States have already fought several times.
As early as April 2016, CanyonBridge, a Chinese private equity investment company, proposed to acquire Lattice. On November 3 of the same year, Lattice accepted the tender offer, announcing that it would be purchased by CanyonBridge for $ 130 million in all outstanding shares, including its own debt.
Lattice is one of the top four FPGS in the world. The other three are Silinx, Altera (acquired by Intel), and Actel (acquired by Microsemi).
Headquartered in Portland, Oregon, USA, it mainly produces communication chips used in cars, computers, mobile devices and other equipment. It can also be used for military communications. It is one of the few manufacturers capable of manufacturing programmable logic chips. one.
This is a business that is willing to buy and sell, but was considered a threat to the national security of the United States by the Committee on Foreign Investment of the United States and could not pass the review. On September 14, 2017, Trump directly issued an administrative order to suspend the transaction. Attempts by Chinese institutions to acquire FPGA technology through acquisitions have failed.
Another important event was that in February 2015, the National Development and Reform Commission of China announced a fine of RMB 6.088 billion for Qualcomm; in March 2017, the United States fined ZTE $ 892 million, which was also close to RMB 6 billion.
Prior to the ZTE incident in April 2018, China's Ministry of Commerce continued to delay approval of Qualcomm's acquisition of NXP, also seeking more protection for Chinese companies. The fierce chip battle between China and the United States has also made people realize the pain of China's chips.
After ZTE's closure, a large-scale discussion was triggered in China. At this time, an AI chip acquisition incident attracted large-scale attention. On April 20, Alibaba Group wholly-owned acquired mainland China's autonomous embedded CPU. IP Core-Zhongtian Microsystems Co., Ltd.
On April 19, Alibaba Dharma Institute announced that it is developing a neural network chip-Ali-NPU. At this point, Ali has made a foray into the chip field through independent research and development, acquisitions and other means.
In the past two years, artificial intelligence has begun to rise globally, especially in China and the United States. The rise of artificial intelligence technology benefits from the increase in computing power. GPUs, NPUs and other chips are indispensable, which has greatly accelerated the iteration and research and development of artificial intelligence products.
In this chip battle, will the war spread to the artificial intelligence industry, especially the medical artificial intelligence industry that we are concerned about. If Sino-US trade frictions escalate again, will medical artificial intelligence companies be "stuck in the neck" like ZTE? Arterial Network has consulted many industry experts on this question.
丨 Medical AI industry's chips rely on the United States
Because the research of medical artificial intelligence needs to calculate large-scale data, the demand for chips is very large.
In the field of medical artificial intelligence, there are mainly two scenarios for using chips to perform data calculations. One is to use the chip to calculate data during the training of laboratory models and the iterative AI products. The other is to embed the chip into an imaging workstation or medical equipment after the product is developed.
There are currently four main AI acceleration chips: CPU, GPU, FPGA, ASIC (Google TPU).
Among them, GPU is the most widely used general-purpose chip, mainly due to NVIDIA's vigorous promotion, high efficiency and reasonable price. The efficiency of the CPU is too low, while ASICs and FPGAs are custom and semi-custom chips. Although the efficiency is high, the low demand results in low productivity and high prices.
In the model training phase, companies use more general-purpose chips. In this process, it will be performed through NVIDIA GPUs, Intel CPUs, and FPGAs from other companies.
At the application stage, the implementation of medical AI products also has a strong dependence on chips. One is a cloud-based solution. Placing the product in the cloud to provide medical services to customers can reduce costs, but the real-time performance is slightly insufficient.
Another way is to integrate AI algorithms and software systems into custom chips, and load custom chips into medical devices to reduce power consumption, ensure system performance, and reduce device size.
This shows that both the model training and the scene application have high requirements on the chip. At this stage, although Chinese medical AI companies are no different from European and American countries in terms of software and algorithms, almost all the chips used depend on US imports.
There are not many domestic companies with AI chip design capabilities. The Cambrian is a representative one, but its packaging production technology is still lacking.
丨 Will the US ban the export of AI chips to China?
At present, almost all chips used in AI products rely on imports. Whether it is Nvidia GPU, Intel CPU, or Google's TPU, the exporting country is the United States.
After the ZTE incident, will medical AI companies encounter the same thing? Industry experts told Arterial Network reporters that the probability is very low. There are three reasons:
* 、 ZTE violates the laws of the country where we do business
ZTE was banned because ZTE made false statements to the Bureau of Industry and Security in 2016 and 2017.
On April 16, 2018, the U.S. Department of Commerce's Industry and Security Bureau used ZTE to fail to deduct bonuses and disciplinary letters in time for certain employees involved in historical export control violations, and on November 30, 2016 and July 2017 Two letters submitted to the US government on May 20 made false statements about this and made a decision to activate the rejection order against ZTE and ZTE Kangxun.
Unlike ZTE, domestic AI companies only use chips for product research and development and domestic operations, and have not yet involved export sales. In particular, the medical AI industry is still in the research and development stage and has not achieved commercial application on a large scale, which does not constitute a threat in itself.
Second, China's chip purchases are huge, and it is impossible to completely block
According to Gartner survey data, in 2017, three of the top ten companies worldwide in purchasing chips were Chinese companies, namely Lenovo, Huawei and BBK, ranking 4th to 6th in the world, and ZTE did not enter the list.
In 2017, Lenovo purchased USD 14.671 billion, Huawei purchased USD 14.259 billion, and BBK purchased USD 12.103 billion.
In contrast, ZTE's chip purchases are not * especially, especially in large companies such as Broadcom and Qualcomm, which are not too high.
According to a Wall Street Journal report, 65% of Qualcomm ’s $ 22.3 billion in revenue in fiscal 2017 came from China, compared with 57% in fiscal 2016.
Of Broadcom's 2017 revenue, 54% came from the Chinese market.
From these data, it can be seen that blocking ZTE will not hurt the backbone of large American companies, but rather that China will purchase a large amount of US chips. Once the chip export is completely banned, it will have a huge impact on the US semiconductor industry.
Third, the overlap between communication chips and AI chips is not high
The chips banned from selling to ZTE in the US this time do not overlap with chips used by AI.
The main chips used by AI companies are GPUs, CPUs, FPGAs, etc. Although these are also imported from the United States, the types of chips used by AI companies and ZTE do not overlap.
Therefore, in terms of the scope of the chip ban, the impact on AI companies is almost zero.
丨 FPGA will be the focus of chip development in the AI field
Although it is not possible for the United States to completely ban the sale of chips, the core technology is subject to others. In this "ZTE incident", people in China have felt extremely ashamed and helpless. During this time, we also learned that independent research and development of chips has become an important trend for the next development of medical AI companies.
In addition to chip design companies, medical AI companies including Inferential Technology and Heath ’s Heterogeneity have also begun to get involved in chip research.
According to the "Comprehensive Interpretation of the Development of the Artificial Intelligence Industry in China and the United States" released by the Tencent Research Institute in 2017, from the number of chip companies at the basic layer, there are 33 in the United States and 12 in China.
In the United States, there are not only technology giants such as Google, Intel, IBM, but also chip giants such as Qualcomm, Nvidia, AMD, Xilinx, as well as many well-developed medium-sized companies and active startups.
Among the chips used by medical AI companies, GPUs and CPUs are mainly used for deep learning algorithms in AI. Regardless of chip architecture, patents, or ecology, these two areas are firmly controlled by NVIDIA and Intel, and there is no chance of getting involved. It can be said that the United States is completely monopolized.
Chinese AI companies want to make a difference, only in the areas of FPGA, ASIC and brain-like chips, and currently mainly small and medium-sized companies.
In the field of AI, the most likely breakthrough for future chip development is FPGA chips.
The Chinese full name of FPGA is called field effect programmable logic gate array, which is a kind of semi-custom circuit in the field of application-specific integrated circuits. Disadvantages. Features high performance, low energy consumption, high flexibility, and hardware programming.
If the CPU and GPU are "generic" at the architecture level, then the FPGA is "generic" at the lower circuit level.
After programming the FPGA through the hardware description language, it can simulate the architecture of any kind of chip, including the architecture of the CPU and GPU.
Baidu's machine learning hardware system is to create an AI proprietary chip using FPGA, and made an AI proprietary chip version of Baidu Brain-FPGA version of Baidu Brain, and then gradually applied it to large-scale deployment of Baidu products, including voice recognition, advertising CTR estimation models, etc.
Chen Hui, CEO of Yasen Technology, introduced to ArtNet that FPGA has eliminated the concept of memory, and a large amount of information is transmitted very efficiently. It can be directly transferred from one unit to another without caching in the main memory, which is especially suitable for real-time requirements. Relatively high algorithm.
Therefore, the integration of FPGA and AI algorithms will become the main research and development direction in the future. At present, there are many companies in China such as Shenjian Technology, Shenzhen Ziguang, Shanghai Anlu, and Horizon Robotics engaged in FPGA-related research.
It is worth noting that the industry barriers in the FPGA field are very high, and nearly 9,000 patents have built long intellectual property barriers. Stronger than Intel is also sighing, not to buy Altera for FPGA tickets for a cost of $ 16.7 billion.
According to Chen Kuan, CEO of Inferred Technology, if China takes the national effort to conduct chip research, it can make high-quality chips, but the industrial barriers to chips are very high. Long research.
Since Intel was founded in 1968, it has taken many years of development to have today's achievements, and China still has a lot of shortcomings in chip industry production.
Ding Xiaowei, founder of Voxel Technology, said that the chip gap between China and the United States has existed for a long time, and the industry is constantly discussing it.
However, although there is a gap between China and the United States in the field of chips, Chinese medical AI companies have their own core technologies in application and AI technology. China and the United States also have their own advantages, but their work processes and methodologies have their own characteristics, and there are differences in defining medical issues and integrating with the clinic.
The understanding of the ZTE incident must be rational. The development of chips needs constant trial and error, a long-term accumulation, and we need to improve the research and development environment and continue to invest.
丨 China's chip development requires joint efforts of BAT and startups
The pain of Chinese chips has made us all pay more and more attention to core technologies. The chip industry has very high technical barriers, long R & D cycles, and very high requirements for funding and teams.
A few days ago, Ali bought Zhongtian Micro's news screen. Many people think that large companies such as BAT and Huawei are rich in funds and attach great importance to AI development. This view makes some sense, but the development of China's AI chips requires not only leading companies such as BAT and Huawei, but also startups.
On the one hand, BAT has strong R & D strength and demand, and they actually have a layout.
In March 2017, Tencent Cloud announced that it has formed a full-matrix AI infrastructure computing platform for FPGA, GPU and 25G network card cloud servers, and also announced a series of technology and ecological deployments, including the launch of 1-machine 4-card FPGA cloud server, 1-machine 8 The GPU cloud server of the card and the 25G network card will be deployed on the FPGA and GPU cloud servers to provide network infrastructure for subsequent GPU clusters and FPGA clusters.
In August 2017, Baidu cooperated with Xilinx at the Hot Chips Conference in the United States to release the XPU, a 256-core, FPGA-based cloud computing acceleration chip.
Baidu also released the DuerOS smart chip. However, this chip is integrated by Ziguang Zhanrui RDA5981 and uses the ARM company's mbed OS core and its secure network protocol stack.
The investment of BAT will solve part of the difficulties of China's AI chips in the future.
On the other hand, we cannot ignore the strength of startup companies. BAT often lays out by acquiring startup companies.
The team of startups in the field of AI chips will be more professional and have a deeper understanding, such as companies such as the Cambrian and Ziguang Guoxin.
The Cambrian is a well-known domestic AI chip research and development company. It is the world's * most successful AI chip company with mature products. It has two product lines: terminal AI processor IP and cloud high-performance AI chips.
The Cambricon-1A processor (Cambricon-1A) released in 2016 is the world's first commercial deep learning processor.
The Cambrian team originates from the Institute of Computing Technology of the Chinese Academy of Sciences, which is China's * national academic institution specializing in comprehensive research on calculator science and technology.
A number of high-tech companies such as Lenovo and Shuguang were born from the institute. It is also an important shareholder of Cambrian technology and a long-term partner of industry, academia and research.
In August 2017, Cambrian Technology completed a USD 100 million Series A financing, jointly invested by SDIC Ventures, Alibaba Ventures, Lenovo Ventures, Guoke Investment, Zhongke Turing, Yuanhe Origin, and Yongli Investment. After this round of financing, the company ranks among the unicorns.
Arterial Network interviewed a number of medical AI institutions. When they chose custom AI chips, they first considered Cambrian products.
In the FPGA industry, "National Team" Shenzhen Ziguang Tongchuang also has a deeper layout. The company was established in 2013 and is a subsidiary of the listed company Ziguang Guoxin Co., Ltd.
In November 2017, Ziguang Guoxin increased the capital of Ziguang Guochuang and established the Chengdu R & D Center with a total investment of about 597 million yuan to promote the development of FPGA and corresponding EDA tools. Ziguang Tongchuang announced on the homepage that it has made China's * 10 million gate-class high-performance independent property rights FPGA.
Another startup, Shanghai Anlu Technology, has a core team from Lattice, one of the FPGA giants. At present, Shanghai Anlu Technology has completed Series C financing. Investors include Huada Semiconductor, Hangzhou Silan Micro, and the Shanghai Municipal Government's integrated circuit fund "Shanghai Technology Venture Capital Co., Ltd.".
It can be said that BAT has funds and needs, and startup companies have technology, understand the industry, and experience, and the cooperation between the two sides will accelerate the development of the chip industry.
* Thanks to Ding Xiaowei, CEO of Voxel Technology, Ding Xiaocheng, CTO of Wofang Technology, Chen Hui, CEO of Yasen Technology, Song Jie, Hei ’s Heterogeneous CEO, Chen Kuan, CEO of Inferior Technology, and Li Yiming, CTO of Shenrui Medical, for their support.
Iridium urine and urine intelligent nursing robot product introduction
"Re-injection" of non-public health care under medical investment
General Administration of Market Supervision: Don't promote health foods for disease prevention or treatment
National Health Commission notified Roche diagnosis, recalled 2,500 boxes of products
County hospital upgraded to third-level hospital! 12 provincial policies are here