AI chip as a carrier rises in all four sects to compete for the 100 billion market

Looking back at the Chinese market, artificial intelligence and AI chip technology have opened a new door for the Chinese market to catch up with the chip industry in the past few decades, making China the first time to stand proud of the world's cutting-edge technology and even lead the global technology trend. Opportunity window.

In this market, capital is the fastest response. Under the catalysis of hundreds of millions of dollars in financing, the AI ​​chip startup market in China is particularly prosperous. In addition to the emergence of financing unicorns, major startups have formed four major factions according to their own characteristics. We can use martial arts. The sects come to be the metaphor of the image - "Shaolin, Wudang, Wuyue, Mingjiao". The players of these four factions have their own characteristics, and the magical powers are fully displayed.

In comparison, the major overseas giants do not give up. Just about 20 days ago, when the whole of China was immersed in the reunion atmosphere of the Lunar New Year, Google’s technology giant Google suddenly announced that its powerful computing power of the AI ​​chip “TPU” would be open to the public – the TPU’s Not too small, Google's AI program AlphaGo (Alpha Dog) is based on the powerful computing power it provides to defeat the world's first master of Go, Ke Jie.

On the same day, Amazon, the technology giant that has been widely known as an online bookstore, was also exposed to a dedicated AI chip for its explosive product, the smart speaker Echo. The project's R&D team has 449 employees.

After half a year, Zhizhi went deep into the industry and reported on the nearly 100 core enterprises in the entire industry chain of AI chips for the first time, covering major domestic and foreign giant players, emerging ventures, scene applications, OEM production, etc. Development, innovation and entrepreneurship have been followed up, which has also contributed to the first domestic AI chip innovation summit held by Zhizhi. This is one of the heavy reports of the intellectual and AI chip industry series. Through the disassembly of the four hottest application industries of the AI ​​chip and the players of the major admissions, the AI ​​chip industry chain has been deconstructed in a panoramic way.

Four major business scenarios face subversion

The significance of the AI ​​chip to artificial intelligence can be understood as the engine to the car. The theory of artificial intelligence has been proposed for many years, but the amount of calculation required to achieve it is too great. This "sports car" has not been equipped with a suitable "engine" and can only be placed in the warehouse. Until the advent of the AI ​​algorithm and the era of big data, and the emergence of AI chips.

Broadly speaking, chips that can drive AI programs can be called AI chips. However, this article specifically refers to chips that have been specially designed for the AI ​​algorithm. According to the application scenario, the AI ​​chip can be simply divided into a cloud AI chip for a cloud server room and the like, and a terminal AI chip for the end smart device and the IoT device.

The cloud AI chip is characterized by powerful performance, simultaneous support for a large number of operations, and the flexibility to support different AI applications such as pictures, voice, and video. The various Internet AI capabilities we use today (such as online translation, human-to-person comparison), behind the cloud AI chip is playing a role or providing computing power.

The AI ​​chip needs to be embedded inside the device, so that the device can have AI capability without networking. They are characterized by small size, low power consumption, and the performance does not need to be particularly powerful, usually only need to support one or two AI capabilities. Now the chips in the mobile phone, the chips in the camera, and even the chips in your home cooking rice are beginning to be AI.

NVIDIA founder Huang Renxun once told wisdom that in the future, AI and AI chips will be everywhere: coffee machines, mugs, microphones, even earrings, shoes and other small items will be intelligent. (Dialogue | Huang Renxun: 5 years revolutionary robot industry AI chip will be everywhere)

Currently, AI chips have four of the hottest business applications—home/consumer electronics, security surveillance, autonomous vehicles, and cloud computing.

1) Home / Consumer Electronics - quietly invade your home, moist and quiet

Let's start with the home/consumer electronics that are closest to our lives.

In September 2017, Huawei released the world's first mobile phone AI chip Kirin 970, which started the first shot of the AI ​​chip invading the mobile phone. In October, Huawei Mate10 and Mate10Pro equipped with this AI chip were officially launched.

This Kirin 970AI chip is equipped with a module dedicated to processing AI - NPU (Neural Network Processing Unit), the technology is derived from the domestic AI chip creation enterprise Sino-Korean. The NPU is 25 times faster than the CPU and will increase energy efficiency by a factor of 50.

This AI chip not only enables your phone to take better pictures, smoother translation, more accurate speech recognition, but also learn to understand your usage habits, let the phone automatically release memory, faster and smoother.

In a series of exclusive in-depth interviews with Huawei Mate10 and Kirin 970 chips, Eric, director of Huawei's wireless terminal chip business unit, once told Zhizhi that the future AI will be a basic technology in the chip, possibly next year (2018). Every chip company will have this ability. (18 months, Huawei AI mobile phone Nirvana birth record)

As it turns out, he is right.

Just two weeks after the launch of Kirin 970, on September 13, Apple released the 10th anniversary iPhoneX, equipped with self-developed AI chip A11, this chip not only makes iPhoneX use faster and smoother, but also makes iPhoneX Support face recognition unlock (FaceID), face recognition payment, photo automatic classification, and real-time expression tracking Animoji and other functions.

On the day of the Spring Festival this year (February 14th), the old chip company ARM officially announced the launch of two AI chip architectures for mobile: object detection processor and machine learning processor. The importance of this matter must not be underestimated - know that more than 90% of the world's mobile phone chips use ARM architecture, even Kirin 970 and Apple A11 are no exception.

At the just-concluded MWC2018, mid-end mobile phone chip giant MediaTek released the new mobile phone chip HelioP60, which supports AI and computer vision, and can provide accurate face recognition and other functions.

The high-end mobile phone chip giant Qualcomm announced on February 22 this year that it will launch an artificial intelligence engine (AIEngine) based on its Opteron chip series, and package all software and hardware AI computing power in Qualcomm mobile phone SoC into this engine for manual use. Smart application on mobile phones is faster and more efficient.

This series of new products announced that the mobile phone has become an important part of the Red Sea battle in the AI ​​chip market. However, in addition to mobile phones, many of the home electronics you use every day are quietly carrying out AI upgrades, mainly around voice AI, representing players including Amazon to create Echo smart speaker AI chips, and Kai Tyron's home appliances. Voice AI chip, Hangzhou Guoxin to create a voice AI chip and so on.

The deputy general manager of MediaTek and the general manager of the intelligent equipment business group said to the East, the global smart speaker market is expected to exceed 60 million in 2018, and in this 60 million, there will be more and more speaker products equipped with AI. chip. (Ten sentence pre-judgment 2018 intelligent voice industry to compete for 60 million smart speaker market)

2) Security camera - a big fire market that is optimistic for all AI chip players

With the rapid development of the market economy, the continuous maturity of technology, and the promotion of national policies, the scale of China's security industry is also growing. According to Jiadu Technology's "Artificial Intelligence Technology White Skin", the size of the security market in 2017 will exceed 635 billion, an increase of 17.6%.

While the scale of the security market continues to grow, it means that the number of security devices based on cameras has increased (in the past year, the shipment of domestic security HD cameras will be around 100 million), and human eye monitoring has not been seen; Coupled with the improvement of security requirements in the society, the security industry pays more and more attention to pre-warning. The traditional manual review method is far from enough to meet the industrial needs. The security requires independent real-time monitoring and real-time alarm.

Because AI can quickly process video and quickly identify people, cars and objects, the machine can not only recognize fugitives and suspects, but also record and analyze his real-time location. Such capabilities are related to public security and traffic police. And civil defense and other needs coincide.

It is no exaggeration to say that almost all AI chip startups now use security as one of the core application scenarios, and they have launched AI chips embedded in security surveillance cameras, and hundreds of millions of R&D funds are directed to this block. "When the red fried chicken" is not an exaggeration.

In addition, several security giants are also eager to move. Security is a traditional industry with a high degree of concentration. The three major industry giants, Hikvision, Dahua, and Yushi Technology, have accounted for half of the industry. These security giants have accumulated rich industry experience, not only the partners of many AI chip companies, but also promote the pace of AI+ security.

For example, Yushi Technology has released a complete set of AI solutions in the AI+ security field, involving front-end smart cameras embedded with GPU chips, face recognition speed-pass doors, and back-end data center all-in-ones. However, Yuhua, the chief architect of Yushi, mentioned to Zhizhi that, unlike AI startups, traditional security companies such as Yushi need to focus on the engineering and landing of this technology in addition to the development of AI technology, such as user equipment room consumption. Engineering problems such as power and heat. The comprehensive rollout of security AI is not that easy. (Dialogue Yushi Chief Architect Yao Hua: tearing open AI+ security and cruel corner, who eats meat? Who drinks soup?)

In the 2018 “Internet +”, artificial intelligence innovation development and digital economy pilot project support project list of Hainan Development and Reform Commission, Haikang Hikvision’s “Computer Vision AI Chip R&D and Industrialization Project” was on paper.

3) Self-driving cars - the power of the three major AI chips to promote the commercial landing of unmanned vehicles

According to the classification standards of the US Department of Transportation, autonomous vehicles can be divided into five levels: L1-L5, L1 driving assistance, L2 partial automation, L3 conditional automation, L4 high automation, L5 full automation (complete automation means no drivers The car can go by itself), which is also a generally accepted classification method in the industry. Due to the need to perceive the environment and make decisions, the L3-L5 level of autopilot technology is increasingly demanding computing platforms, and the demand for AI chips is growing.

At present, AI chips are becoming a core component of the autonomous driving computing platform. The chip giants such as NVIDIA and Intel, and the startups such as Horizon, have integrated into an autonomous driving computing platform for OEMs and Tier-1 (first-tier supply). The auto parts dealer giant has landed in the autonomous driving solution. There seem to be three forces at the moment:

1. NVIDIA's Xavier computing platform (formerly DrivePX), which is being used by more than 20 autonomous driving startups, as well as supplier giants such as Bosch and ZF, to create their own autonomous driving. The closest to mass production is the ZF ProAI solution.

2, Intel GO computing platform (MobileyeEyeQ chip acquired by Intel CPU, Intel, Altera FPGA processor acquired by Intel) is OEM BMW, Italian FCA, and supplier giants Continental, Canada Magna, Delphi (post split) Out of Anbofu), they have successively adopted it as an autonomous driving solution. Among them, Intel, BMW, and mainland China also led the establishment of the Autopilot Alliance.

3. The startup company represented by the horizon is building its own autonomous driving computing platform. The Horizon's Hugo Autopilot platform used the Intel FPGA processor in the early days, and later built the self-developed BUP architecture, and introduced the AI ​​chip "journey" that follows this architecture. In addition, Zhizhi learned that due to the huge auto-driving industry, many AI chip startups in the first camp also secretly targeted this field.

Regardless of the above-mentioned autopilot computing platform, the AI ​​chip is a key component and the main driving force for the development of autonomous driving technology.

Bosch, as the leading player of Tier-1 in the car, will realize L2 level automatic driving in 2018 and L3 commercial in 2021. Bosch Chassis Control Systems China's driver assistance system radar research and development department director Cai Wei told Zhizhi that Bosch is cooperating with many autopilot computing platform suppliers, including giants and start-up companies. Bosch has high requirements on power consumption and cost for autonomous driving computing platforms.

For computing platforms that require automated driving, Bosch is collaborating with a number of suppliers, both giants and start-ups. Bosch has high requirements on power consumption and cost for the autopilot computing platform. In the future, Bosch's choice will be an embedded computing platform.

Zhang Dezhao, CEO of Automated Driving Startups, has also told wisdom that autopilot technology will take time to spread, so startups must focus on technology and commercial realization while conducting technology research and development. And while unmanned vehicles require a strong computing power, cost, power consumption and mass production are all elements to consider. (Take the unmanned car of the CEO of Jingdong Xiaomi Investment Zhixing to earn money)

4) Cloud computing - powering the Internet AI, efficient and flexible

Regardless of Weibo or point take-out, we now use all Internet applications, behind which cloud computing data rooms provide computing power. And the various Internet AI capabilities we use now (such as online translation, human-to-person comparison, image search, etc.), behind the cloud AI chip in the huge data center room around the clock to provide computing power. .

As mentioned above, the cloud AI chip is characterized by powerful performance, ability to support a large number of operations at the same time, and flexible support for different AI applications such as pictures, voice, and video. This is a huge market, and it is also the tightest battlefield for all kinds of chip giants.

The first shot was Nvidia, the young chip company founded in 1993 in the artificial intelligence era, because its GPU is particularly suitable for today's mainstream AI algorithm - deep learning - training, this home Not only did the company's share price soar from more than 30 US dollars to more than 200 US dollars, it also set off a massive AI chip boom around the world.

The first response was the world's major traditional chip giants, Intel, Qualcomm, ARM, MediaTek, etc. have come into play. Among them, Intel, which has dominated the chip market for more than a decade, is a big gold out of the warehouse, all the way to "buy and buy", layout of many AI chip product lines, cloud AI chips and end AI chips are involved.

Intel Institute Dean Song Jiqiang once told Zhizhi that the future of AI chips must be diversified, and different types of products meet different power consumption, size, and price requirements. Intel’s current AI chip layout is moving toward this diversity. development of. Artificial intelligence is a marathon, and this game is just beginning. (Dialogue Intel AI three executives: optimistic about the diversity of AI chip introgressors constantly adjust AI investment logic)

In addition, domestic technology giants such as Baidu, Ali, Tencent, and Keda Xunfei have also entered the cloud AI chip field, but for the time being investment trends, Baidu released XPU in August 2017, which is a 256-core, FPGA-based cloud computing AI chip.

Four major sects: Shaolin, Wudang, Wuyue, Mingjiao

In terms of cloud AI chips, overseas players represented by NVIDIA GPUs started earlier. In terms of the terminal AI chip, many startups in China are in the forefront of the world.

Professor Bill Dally, an academician of the National Academy of Engineering and chief scientist of Nvidia, once told wisdom that the market for cloud AI chips is now relatively mature, and the world's major technology giants are deeply rooted and the pattern is hard to be shaken. In contrast, the wide variety and huge number of terminal AI chip markets have yet to be expanded, which is the opportunity for many AI chip start-ups. (Exclusive dialogue NVIDIA chief scientist: decoding AI chip battle)

Under the catalysis of hundreds of millions of financing, the AI ​​chip startup market is particularly prosperous. Not only have a number of AI chip startups emerged since 2015, but also many unicorn companies have emerged. According to the characteristics of the founding team background, company attributes, technology genre, etc., these domestic AI chip startup companies can be divided into four major factions: "Shaolin, Wudang, Wuyue, Mingjiao". (Respond to the "AI chip rivers and lakes" in the smart things dialog box to get the AI ​​chip four martial art business information summary table)

1) Shaolin: a strong academic, one-by-one heritage

The reason why these are called “Shaolinists” is because these entrepreneurial teams were generally established earlier, mostly from Tsinghua University, Peking University, Chinese Academy of Sciences, University of Electronic Science and Technology, and other institutions of higher learning/research institutions. They have strong academics. Attributes. There will also be an “old professor” blessing in the founding team. They have many years of academic accumulation in chips and related academic research, and have served in these institutions and institutions for many years, providing a systematic talent resource for the company. System – For high-tech companies, talent means technological advantage and also means strong market competitiveness.

This faction is not only favored by capital with its deep academic background, but also able to obtain support from the “parent school” in terms of market and talent. For example, one of the representative players: Zhongke Cambrian - its founders Chen Yunqi and Chen Tianshi are both doctoral graduates of the Institute of Computing Technology of the Chinese Academy of Sciences. At the beginning of the business in 2016, Cambrian not only obtained the 1000 Institute of Computing Technology of the Chinese Academy of Sciences in the Angel Wheel. The research funding of 10,000 yuan has also won the support of the Chinese Academy of Sciences in various project resources. After the launch of the Huawei HiSilicon 970 chip with the Cambrian AI chip technology, the Institute of Computing Technology of the Chinese Academy of Sciences also sent a congratulatory letter to Huawei, emphasizing the background of the Cambrian Academy of Chinese Academy of Sciences.

In addition, on behalf of the players and Tsinghua Electronics Department, Professor Wang Yu led the team founded by Shen Jian Technology, University of Electronic Science and Technology University doctoral tutor Liu Yang Professor led by Shen Si Chuangxin, Tsinghua University Professor Wei Shaojun led by the Institute of Microelectronics, Tsinghua University, The brain-like computing chip team led by Shi Luping's teacher group and the FPGA-based deep learning acceleration team led by Cong Jingsheng, director of the Center for Energy Efficiency Computing and Application, Peking University.

Among these players, Shenjian Technology is one of the earliest independent enterprises to operate as a commercial enterprise. It was established in March 2016 and was originally started by FPGA technology. It has successfully entered the first camp of China AI chip creation enterprises. Shen Jian, CEO of Shenjian Technology, revealed to Zhizhi that at present, Shenjian has received tens of millions of orders, and both AI chips are in mass production, and many partners are already using them. (Hands of thousands of security orders exclusive secret deep AI chip ace weapons)

Dr. Yu Dejun, CEO of Shensi Chuangxin, told Zhizhi that in addition to the existing folk scenes, AI chips are needed. Military missile aids, ships, submarines and other scenes also need AI chips. Shenxin Chuangxin is currently working with domestic military units. Cooperate with these projects.

The Institute of Microelectronics of Tsinghua University, led by Professor Wei Shaojun, has been working on the development of cutting-edge academic technology for more than 30 years. AI chips belong to the technology branch. Last year, the AI ​​chip Thinker of the Institute of Microelectronics of Tsinghua University was at 2017 ACM/ The IEEEISLPED International Low-Power Electronics and Design Conference won the Design Competition Award, which is the first time that a Chinese mainland unit has won this honor in the first completed unit. (The ultimate problem of AI chip was solved by Tsinghua University IC male god)

2) Wudang: old brand, cross-border transformation

Wudang, the ancestor of the family. In the big rivers and lakes of AI chips, in addition to various startup companies, many traditional chip manufacturers have also entered the market. AI chips are a natural upgrade for their existing products, and a time window where opportunities and challenges coexist. - Well done, overtaking in the corner, becoming the industry leader in one fell swoop; not doing well, it is easy to lose the first opportunity, and even the danger of being eliminated by the market.

These companies have long been known for their traditional set-top box chips, mobile phone chips, and security camera chips. Compared with AI chip startups, these established chip companies generally have rich experience in design and manufacturing, market channels, and mature business operation experience. Representative players include Huawei Hisilicon, Hangzhou Guoxin, Zhongxing Microelectronics and so on.

For example, Hangzhou Guoxin was established in 2001 and is one of the earliest domestic chip design companies. It currently has a market share of 15% in the global set-top box chip market. At the end of 2017, Hangzhou Guoxin also released its first voice AI chip GX8010.

It is worth mentioning that because security is one of the core scenarios of AI chip landing application, and the market is huge. But now almost all security chip giants, security hardware giants (such as Yushi Technology, Guoke Micro, Zhongtianwei, etc.) are gradually carrying out research and development and cooperation related to AI chips. At present, there are not many news published by these manufacturers, but they are all in secret.

3) Wuyue: Shentong shows up, Haidu Eight Immortals

Most of the players of the “Five-Yue School” have accumulated many years of technical experience in the traditional chip industry. Compared with the players of the previous two factions, they have many years of chip-making experience and have deep understanding of security, home and other industries. It also has the market attributes of a fast-paced decision-making and fast running of a startup company.

However, due to the large amount of R&D investment in the early stage of the chip, the players of the “Five-Yue School” generally have investment figures of technology giants, such as Baidu, Tencent, Ali, and Intel; and such as Yitu, Yunzhisheng, Yuntian Lifei, etc. An artificial intelligence startup with a certain amount of capital.

The representative players of this group include Intel's investment in the horizon robot. It is worth mentioning that Yu Kai, the founder of Horizon Robot, is the former deputy dean of Baidu. In addition, there are also investment companies such as ThinkForce, Alibaba Entrepreneur Fund and Qualcomm, Kneron, Yunzhixin, Yunzhixin, Roobo, Kaiying, and Shanghai Baoxin Software. Haiqing Zhiying and so on signed the cooperation agreement.

4) Mingjiao: Bit blessing, another way

Among the many AI chip companies, there is also an alternative player: they have achieved annual revenues of $2.5 billion in just a few decades of entrepreneurial time, and the accumulation of wealth is almost horrific; they worship wildly. The technology, as much as possible, squeezes out the last calculation performance of the chip; they are blindfolded and thrown away, but at the same time they are mysterious and low-key, and they don't say anything about the suspicions of the outside world.

They are the water sellers in this wave of bitcoin frenzy - the producers of the virtual currency "mine machine chips", representing players including Bitland, Jianan Zhizhi and so on.

The acquisition of virtual currency requires "mining", and the essence of "mining machine" is a kind of mining computer, which is also powered by chip drives. The more powerful the chip is, the faster, more and more advantageous it is to dig up the mine. In markets where virtual currency prices are skyrocketing, the wealth of these chip companies that provide engines for miners has naturally risen.

In terms of Bitumen, Bitcoin was established in October 2013. Its total revenue in 2017 has soared to US$2.5 billion. It has more than 1,000 employees and has established a computing center with tens of billions of calculations worldwide. The chip market has more than 70% share and has an overwhelming market advantage.

However, at the end of 2017, when the virtual currency skyrocketed, Bitcoin took a different approach and officially launched the first AI-specific chip “SOPHON” – the name comes from the joint CEO of the Bitland Continental, Jank. The set of science fiction novels, "Three-body."

Bitumin's product strategy director Tang Weiwei also revealed to Zhizhi that in the past two months, Bitian's AI chip team has expanded to 300 people.

The chip has always been a "burning money" thing, this 28-meter process "Yuanfeng" AI chip only costs millions of dollars for development and filming. The iterative speed of traditional chips is 1-2 years, but Bitcoin announced that its products will be rapidly iteratively developed at the speed of 9 months. (Exclusive: See how the mining industry has turned the sub-bital continent to the AI ​​chip)

In addition, Jia Nan, the world's second-largest bitcoin mining machine manufacturer, also pre-released its self-developed AI chip KPU in December 2017. At present, the mining giant has applied for the listing of the new three board. In 2017, the company's revenue exceeded 1.2 billion yuan and its profit exceeded 300 million.

Compared with other companies that started cutting into chips from AI, the mining machine has rich experience in chip creation and rich capital accumulation. However, other startups focused on AI chips are more AI algorithm optimization and software ecology. Have an advantage. After all, for AI chips, both AI and chip capabilities are important.

At the beginning of the source: cloud computing, the first batch of overseas giants who ate meat

The so-called artificial intelligence is actually to make the machine have the same intelligence as people. But computers have always been very "stupid". Some tasks that are easy for humans are difficult for the computer to climb into the sky: for example, what is the animal in this photo?

For humans, we saw a gray cat sitting in the wild, and for the computer, it saw a set of numbers in the image; the cat ears and cat eyes we saw, in the eyes of the computer Just the difference between numbers 88 and 23.

An important discipline derived from this is called ComputerVision (CV), and it is clear that the two similarly-shaped animals, "cat" and "dog", and the position of the cat in the picture have been The subject of this discipline is at the Pearl level, and it has been difficult to make progress in the past 20 years.

Until 2012.

In the early winter of 2012, a big thing happened. In the 2012 Olympics "Olympic" - ImageNet Challenge, there was an alternative player, an old professor from the University of Toronto, Geoffrey Hinton. Hinton and his team used the GPU for the first time. Chips and deep learning algorithms have dramatically reduced the error rate of computer graphics by several times, becoming an important node in the history of computer vision.

In the 2015 ImageNet contest, the Microsoft Research Asia team relied on GPU and deep learning algorithms to make the computer's ability to map more than humans for the first time. The human error rate is about 4%, and the error rate of the champion team's machine map is 3.57%.

A stone stirred up a thousand waves. GPU chips and deep learning algorithms have become the "net red" of artificial intelligence, and countless scientists have begun to shift their research focus to them.

A deep learning algorithm can be understood as a computer program that implements artificial intelligence. It allows the computer to learn the characteristics of "cat" by learning a large amount of material (such as thousands of different kinds of "cat" pictures). Next time you give it a photo of the cat, whether it is a black cat, a white cat or a Persian cat, it can be recognized.

After the emergence of image classification and recognition, a series of AI applications such as video recognition, speech recognition, translation, and voice assistant have emerged.

In fact, the GPU's life is more widely known - graphics card. In the wrong place, the scientists found that the parallel computing architecture of the GPU chip and the deep learning hit it off, suddenly compressing the "learning" time that the machine originally took for several months to a few days or even hours, the two strong partners, In 2012, ImageNet shined.

However, Intel, IBM and other established cloud server chip manufacturers are also actively deploying this market, each of which continues to cut into the cloud AI chip market through mergers and acquisitions, investment, research and development. (Three and a half years! Is Intel's AI chip torpedo going to launch?)

At the same time, as a representative of innovative technology, Google has also started to build its own cloud AI chip - TPU (tensor processing unit) from 2014, Google's AI program AlphaGo (Alpha Dog) is based on its powerful Computational power has defeated Ke Jie, the world's first master of Go. About 20 days ago, Google announced that the powerful computing power of TPU will be open to users through Google Cloud.

Li Jia, president of Google AI China Center and director of R&D of Google Cloud AI, once told Zhizhi that the AI ​​chip boom will affect the AI ​​industry for a long time. The progress of the chip will bring a positive support to the development of the artificial intelligence industry. The progress and development of computing power will bring more opportunities and more imagination to artificial intelligence in the future. (Dialogue Google Li Jia: AI chip boom will continue to promote AI development AutoML users have exceeded 10,000)

Since then, artificial intelligence has ushered in the largest wave of market application since the establishment of the discipline in 60 years. From the once-unreachable cutting-edge technology, it suddenly came to us and became a household name.

If 2015-2016 is the fierce battle for AI chip giants to fight, then the next 2016-2017 is the time when the domestic AI chip startup market is heating up, eventually becoming a big fire, and entering the white-hot competition. As the market heats up further, by the end of 2017, market competition has reached an unprecedented level of intensity.

Just during the two months from October to November 2017, four domestic AI chip companies announced that they would receive large amounts of financing of more than 10 million US dollars. If they start from August 2017, they will be four months. At least a dozen AI chips are available. The upgrade period of chip products is generally 12-24 months. Nowadays, such a hot and hot market bombardment is like a collective carnival of AI chips.

Thinking after carnival

However, behind the collective carnival, there is always a calm voice. Many people in the industry have expressed their concerns about the AI ​​chip industry to the intellectuals:

1) The chip competition is very fierce and it is a "one of the best" industry. The advantage of the chip lies in its scale. The sales volume of a common chip needs to reach a million level to achieve breakeven. Once the market matures, the giants can harvest the market through ten times and a hundred times the output. These large number of emerging AI chip startups may eventually run out of one or two.

2) There are many publicity technologies for artificial intelligence, and there are few industries. Almost all AI chip startups have set their main goals in the security and autonomous driving markets. For security, the market demand for AI has existed for many years, but AI applications in large-scale remote scenarios. There are still many engineering problems that need to be solved. For autonomous driving, one technology has been used for a long time to commercialize on a large scale. Second, the major automakers regard the safety attributes as the highest, and the current black box for artificial intelligence. "The situation is more difficult to accept. Now, apart from mobile phones, there is no AI chip sold more than 1 million.

3) The artificial intelligence algorithm is still in the process of rapid development, and the calculation model is changing every six months to one year. The design and development cycle of the chip is generally long, and only very mature algorithms are suitable for curing onto the chip. The artificial intelligence algorithm is not mature enough.

Professor Wei Shaojun, chairman of the IC Design Branch of China Semiconductor Industry Association and director of the Institute of Microelectronics at Tsinghua University, believes that from the perspective of industrial development, AI chips will continue to be hot in the next two years, and everyone will get together; but by 2020 Before and after, there will be a group of exiters, and the industry will begin to shuffle.

Chip, the root of the country

As the "brain" of all electronic products, the importance of chips for the development of technology need not be rumored. For a long time, the chip technology in mainland China has been lagging behind and relies heavily on overseas imports. Since 2013, the amount of chips used in mainland China has exceeded 200 billion US dollars. The chip has surpassed oil and become the largest imported commodity in China.

Compared with developed countries and regions, the gap in chip manufacturing technology in mainland China is not small. Although the domestic chip design industry has maintained rapid growth (about 20%) in recent years, in general, industrial development still has problems such as weak independent innovation capability, external dependence of key core technologies, and lack of talents. For example, TSMC, Intel, etc. are already developing technology for 7nm chips, and the 28nm process in China's mainland is just beginning to spread smoothly. The 14nm process is still under research.

However, the rise of AI chips has opened a new door for China's chip industry, which has long been able to catch up. On the battlefield of this kind of emerging technology, we stood on the same starting line with the developed countries for the first time, and even stood at the forefront of global technological development in some areas.

The state has paid more attention to artificial intelligence and AI chips than ever before. In July and December last year, the State Council and the Ministry of Industry and Information Technology successively released the "New Generation Artificial Intelligence Development Plan" and "Promoting a New Generation of Artificial Intelligence Industry". The Development Three-Year Action Plan provides directional views on the next development plan of the artificial intelligence industry. In addition to China's new-generation artificial intelligence "three-step" strategy for 2030, the neural network chip (AI chip) is also mentioned as one of the three core technologies.

The Action Plan states that by 2020, domestic AI chip technology needs to make breakthroughs, including cloud neural network chips and terminal neural network chips.与此同时，AI芯片需要实现在智能终端、自动驾驶、智能安防、智能家居等重点领域的规模化商用。

而2018年政府工作报告中还提出，要加强新一代人工智能研发应用，在医疗、养老、教育、文化、体育等多领域推进“互联网+”。

结语：重塑世界的AI芯片

在一场AI芯片大火烧过了2017年之后，2018年，我们将看到一大批AI芯片初创企业的产品正式落地商用，各大芯片巨头玩家的相关产品也会逐一面市，我们即将进入“无芯片不AI、无终端不AI、无行业不AI”的时代当中。

未来，每个芯片都提升它的AI计算能力，小到一个耳机、一台音箱，大到一台工业机器人、一辆自动驾驶汽车，这些终端的AI化将深入到各行各业当中，除了文中重点解读的四大行业，未来，工业、制造、医疗、教育等行业也会逐渐被AI芯片渗透入侵;随着AI计算逐步渗透社会各行各业，我们的世界将逐渐被AI重塑，我们也将进入一个AI社会。

随着巨头玩家的不断入局、创业公司的快速奔跑，AI芯片——作为人工智能产业皇冠上的明珠——已经逐渐成为了一场高手间的较量。这一新兴技术既为科技巨头带来了业务升级、产业扩大的新风口;又也为各大创业者提供了颠覆现有格局，重塑科技话语权的崭新机会;同时还为中国半导体产业实现超车提供了一个绝好的新机遇。