“Are you going to make AI chips?”
Maybe this is a problem that many AI companies will encounter.
According to IT Orange’s statistics, in 2017, the number of domestic AI investment projects reached as high as 384 and the total investment exceeded 62.2 billion yuan. It is noteworthy that, of these, a total of 139 companies in the field of computer vision to obtain financing, the total investment has reached 22.5 billion yuan. For example, one after another last year, refreshing the financing record Shang Tang, Kuangtui.
At the same time, the high financing behind also revealed a new trend: the technology company is sinking, began to turn its attention to the underlying layer of chip development. Just in 2017, a rough estimate there are dozens of startups announced to do AI chip.
In addition, some data show that in 2016, the global market size of AI chips is 2.388 billion U.S. dollars and it is estimated that the global market size of AI chips will reach 14.616 billion U.S. dollars by 2020.
Behind the Rise of the AI Terminal Chip: Edge Computing
The traditional processor is mainly CPU, with the calculation of the increase in demand, GPU AI industry at this stage into a fragrant . This is also a big factor in NVIDIA’s continuous stock price rise in recent years.
The AI chip we are referring to now refers to a chip that is specifically designed to speed up the AI algorithm.
Based on the commonality and differences in computing performance, we can usually divide AI chips into three categories: FPGA, ASIC, brain-like chip and so on.
Computer Vision Inc. Dejected As disclosed in an interview that the company is developing FPGA chips, but according to technology also said that there are plans to do the chip in the future, coincidentally, they voted last year, artificial intelligence chip company ThinkForce. Shang-tung at the end of last year received from Qualcomm’s strategic investment, the stakeholders said that the chip side, Shangtang mainly Qualcomm cooperation.
At present, the main usage scenarios of AI chips can be divided into cloud (server) and terminal two categories. In the cloud, GPU-led NVIDIA, and Intel, Google’s TPU and the mainland’s mainland China also launched their own dedicated chips.
However, with the rise of AI, the demand for computing power is getting higher and higher, and the cloud has more data pressure.
If all data processing is done in the cloud, first of all it puts pressure on the bandwidth and real-time transmission of communications, secondly information security and privacy issues, so high performance and low power terminal intelligence is put on the agenda: Put more Data processing on the device side close to the data source, some of the AI calculation of the pressure transferred from the cloud to the edge of the end.
That’s what we’ve been saying about the edges:
Edge computing refers to an open platform that converges network, computes, stores and applies core capabilities at the edge of the network near the source of data or data and provides edge intelligence services nearby to meet the requirements of industry digitalization in agile connectivity, real-time services, data optimization, application intelligence, Security and privacy protection and other key needs.
This is why a growing number of AI chip manufacturers a major factor.
In addition, “Cloud server chips can not be done by anybody. Some servers are powered by a nuclear power station. Naturally, we can only concentrate on the AI terminal ‘interface chip.'” National “Thousand-Person Project” Distinguished Specialist, Lin Fujiang, vice president of University of Science and Technology of China Microelectronics said.
In this situation, the AI chip suitable for the terminal side equipment also came into being. So, in the traditional chip market has been monopolized by foreign giants now, AI chip will continue to repeat the same mistakes?
The tuyere on the AI chip is a game of foreign giants?
Magnesium guest network, “do not do the chip necklace? Domestic high-end IC chip break has been imperative | Interview with CUHK Vice President, Zhejiang University professor, “a man mentioned:
Although China has digested nearly one-third of the market demand to become the world’s largest consumer of chips, but behind the prosperity there is a cruel fact: China’s domestic chip less than 30% self-sufficiency, the output value of less than 7% of the world, the market share It is less than 10%, that is to say more than 90% of China’s “core” rely on imports.
By the end of 2016, the import amount of China’s chips reached about 1.3 trillion yuan, while the import of crude oil during the same period was less than 0.7 trillion yuan. China’s spending on chip imports is nearly double that of crude oil.
At the same time, security giants such as Hikvision, UOB Shares, and UTS Technologies maintain close ties with chipmakers such as NVIDIA and Intel.
Various cases show that in this piece of semiconductor, the country has lagged behind foreign countries. So AI chip is also a catch-up opportunity, which is why some start-ups will be supported by the background capital of state-owned assets, such as Cambodian investors, including SDIC Ventures and NSRI.
In the meantime, in December last year, in the notice of the Ministry of Industry and Information Technology on the issuance of the “Three-year Action Plan for Promoting the Development of a New Generation of Artificial Intelligence Industry (2018-2020)”, it was mentioned that it should be implemented in intelligent terminals, automatic driving, intelligent security, Smart home and other key areas to achieve large-scale neural network chip business.
AI chip as the most basic part, do a good job in advance “siege slightly pool” preparation, but also pave the way for follow-up development. Data shows that our country has also formulated the goal of developing the semiconductor industry. In 2016, the domestic chip production rate was only 26.2%. By 2025, the domestic product rate will increase to 70%. This means that the domestic semiconductor manufacturing capacity should also increase simultaneously.
In contrast, AI chip research and development, the current foreign semiconductor giant action is not large, mainly to acquisitions and cooperation-based, Intel, for example, they have acquired Altera, Nervana, Movidius, Mobileye and more Home company, won the FPGA and other chip processor technology.
According to Wang Zhongfeng, a distinguished expert from the “Thousand Talents Program” of the country, IEEE Fellow and professor of Nanjing University, “chips are generally giant games, but many large international companies such as Broadcom in the United States and some big ones in Japan and South Korea R & D investment in the AI company is not large, mainly because of machine learning or artificial intelligence algorithm is not the traditional strengths of these companies, of course, some focus on profit margins do not want to enter the not mature AI market. The start-up AI chip company If you focus on a specific application scenario is likely to have a place in the fierce competition in the market. ”
Professor Lin Fujiang think that domestic start-ups to do AI chip is able to break through, “‘interface’ chip (for the terminal chip) of the special applications, the hardware is not too large, and algorithm-based, small companies have a great opportunity of.”
AI chip company “fragmented”, there will not be a dominance
Indeed, most computer vision or NLP start-ups now rely more on technology scenarios and develop the appropriate chips around the terminal side. It is no exaggeration to say that software and hardware solutions + before and after the end all-you-go is becoming a trend.
Yu Kai, the founder of the Horizon, also said in an interview earlier that “the chip was eventually used and not used to send essays, so it depends on how the problem is solved in a specific scenario.” The traditional chip may Regardless of the subsequent application, so they can only do their level, the number of watts how much computing power, and did not consider the computing power of the application of meaning, this is the traditional chip.
AI chip “back to the beginning of heart”, then its essence is for the specific application scenarios and services. Li Xiaoxian, vice president of cloud know the sound loT division that “the technology is actually followed, the most crucial depends on your specific scenario to solve the problem, starting from the specific scenario, to deduce the chip can not solve the actual problem.”
Magnesium guest network access from some AI companies point of view, at this stage the application of AI chips are mainly financial, security, Internet of things, automatic driving and other segments of the main scene.
“There are currently dozens of AI chip companies in the country, and the coincidence degree will not be low, but most companies have their own focus, such as the main processor of the Alps dedicated to the Cambrian era, and the robot developed in the autonomous area Shang Shang has accumulated better intelligence in monitoring. ”
Professor Wang Zhongfeng said: “The final market will inevitably eliminate most of the companies, some will merge with each other, and some companies have been acquired by big companies. They may independently develop and successfully list less than ten.”
But in those segments with a high degree of coincidence, eventually one or two chip makers will not monopolize it?
In response, Professor Lin Fujiang think, AI chip development to the latter will not be such a situation, “AI chip is more of a concept industry, is to introduce some things that can solve the algorithm problem to the chip, I do not think there will be universal AI chip, there will not be one or two big chip makers to unify. ”
AI chip is phased, expect it in a year or two back slightly slightly urgent
Under normal circumstances, the chip development cycle are calculated according to years, according to the release time last year, there will be a lot of AI chips in the second half of 2018 will be available.
Such a long R & D investment may not be synchronized with the development of algorithms and applications. This uncertainty also poses an unknown risk. Especially for the chips burn the hardware, a little flaws, the loss is more than ten million yuan.
In other words, although the domestic chip industry is huge, the overall R & D cost is very high and the success rate is hard to guarantee.
Moreover, it is not enough for the AI chip to enhance the deep learning capability. The algorithms and computational characteristics of each link need to be different because sensor access, signal processing, detection and identification, and decision-making and feedback at the software level are all necessary.
Li Xiao cold clouds that the AI chip has three elements must be considered:
First: There is no corresponding chip knowledge, will not do the chip.
Second: There is no algorithm and application, in fact, from the core point of view, we are looking for a suitable application platform for the algorithm, if we do AI chips, we must have AI algorithm, the algorithm and the hardware is bound.
Third: to have their own business model, when the chip came out, how to sell it, who is your customer, you are ready to put it into what kind of product form, these business models are to be considered.
From this point of view, there are some companies now have algorithmic technology, but if only from a single point of departure to do research and development of chips, in front of them is the follow-up applications and the development of the terminal market.
Li Xiaohan told Mgm.net, “Although it is profitable to do the chip, it is too radical to expect the first chip to come back.”
“The chip is not as good as it is, with a lot of output and low output, and nowadays so many people do the chip, I think most of them will be eliminated.” Yun Sun, co-founder of the technology, said.
AI chip is a long-term investment in the industry, it has its own evolutionary route, it is basically impossible within a year or two, one-stop, and began to make money. For example, when the first chip is introduced, the chip maker must continue to make the appropriate optimization, including adding features, reduce costs, and its production testing also has its own cycle.
As most startups think: The AI chip is a long-term process, expecting it to get in a hurry in a year or two.
An industry source said, “I did not see why everyone started to do the chip, but the chip in this direction is good, it is a market development direction, but not as impressive as the industry, the bubble a bit big in the profit Aspects, the chip itself is not easy, not to mention the AI chip is still speculation concept.
Indeed, combining a chip’s performance, power consumption, and usage scenarios to make a really good product is in fact a huge challenge for every business.
Professor Lin Fujiang believes that only wait for the quantum computer based on spin electronics to become popular 30 years later, there will be a real AI terminal appears.
2018 is to test the beginning of these AI chip makers, in the end what are the start-ups can be small and beautiful or large and in the process of survival of the fittest, the chip market structure will be a big change, even if a company like NVIDIA, At any time is also facing the crisis by the new Black Horse counterattack.
And many do not have too much semiconductor background into the chip a large number of areas, but also to release a signal: rain to wind floor, change the days not far away.