NewsCompany newsIndustry News



With the advent of AI boom, how many paths can we choose?

Chips drive every technological revolution. At the same time, the new era will bring new growth points to the chip industry. Today, the AI boom is driving the rapid growth of the AI chip market. According to a report released last year by Allied Market Research, the global machine learning chip market will be about $2.4 billion in 2017, which is expected to reach about $37.8 billion by 2025, with a compound annual growth rate of 40.8%.

The fast-growing and short-term AI chip market, which will reach tens of billions of dollars in size, not only drives the transformation of traditional chip companies'strategies and technologies, but also attracts many new players, including giants, start-ups, hardware specialists, and algorithms and software entries. But in any case, the key for enterprises is to be able to occupy a place in the market. So, how many paths can chip companies choose?

MetInfo enterprise content manager system | MetInfo CMS

Eco-warfare of giants

At present, there are two kinds of giants in the field of AI chips, one is traditional chip giants, such as Intel, Invidia, AMD, and the other is technology giants that are better at software, such as Google, Facebook, Amazon, Baidu, Ali and Tencent.

Traditional chip giants are doing the right thing

Traditional chip giants tend to enjoy the benefits of the new era earlier. For example, the third AI boom started in 2012. With the parallel computing advantage of GPU, when many entrepreneurs had just begun to prepare for the start-up of AI chips and had not even started their business, Invida launched the first Pascal GPU for in-depth learning optimization in 2016. With the shipment of Pascal GPU, Invida achieved the goal that many AI chips companies want to surpass. Of course, in order to get higher revenue from the AI market, Nvidia is continually updating the GPU for in-depth learning optimization.

Intel is also rapidly benefiting from AI, and an important reason for their benefits is the popularity of cloud computing and the growing demand for AI in data centers. As a result, Xeon (Supreme) chips, which dominate the server market, are the first to bring AI revenue to Intel, which has been around for more than 50 years. However, the problem facing Intel is that CPUs with strong versatility are not good at parallel computing, so it is obviously not enough to maintain the leadership of chips only by strong processors. Strategic and technological transformation is also needed.

So in mid-2016, Brian Krzanich, then Intel's CEO, said on his blog: "We need to transform Intel from a PC company into a company that drives cloud computing and hundreds of millions of intelligent, interconnected computing devices." Today, Intel has resolutely shifted from transistor-centered to data-centered, and has set "interconnection of all things" as its future development trend.

Intel supports strategic transformation through six technical pillars, including process, architecture, memory, interconnection, security and software. In addition to CPU, Intel has already acquired FPGA, autopilot processor and neural network processor through acquisition. Next year, it will launch independent GPU Xe, which can be used in AI.

Technology giants are "forced" to become self-sufficient

Compared with the traditional chip giants in order to better meet the market demand, the technology giants who are better at software have a sense of being forced into the chip field. When the popularity of intelligent devices reaches a certain level and AI is hot, the technology giants who accumulate a large amount of data are more aware of the advantages that mining data value can bring, and the strong demand for AI chips is also generated.

However, with the development of AI algorithm and the application of AI technology, technology giants have found two obvious problems. First, the computing power of existing AI chips can not meet their needs well, and second, the price of GPU for deep learning is high. Therefore, inadequate computing power, high chip prices and personalized needs are not well met as the driving force for technology giants to independently develop AI chips.

Startups have their own unique skills

If the giants are forced to enter the AI chip market, they have at least enough capital and need not worry too much about the acceptance of the chip market. However, AI chip start-ups are not only short of funds, but also faced with great uncertainty about whether they can get market recognition. Therefore, for AI chip start-ups, they need to do their best to share in the AI chip market.

Since resources are limited and risky, start-ups need to think carefully about many issues in addition to understanding their strengths. For example: do you want to do cloud or terminal chip? Do you want to make a visual chip or a voice chip? Which chip type is more suitable? What markets are they facing? What business model do you use? This series of problems indicate that the threshold of the chip field is high enough and the difficulty is big enough. It is more difficult to gain a place in the emerging AI chip market, but there are more paths.

Buy by a giant

For startups, acquisition by giants is clearly a good outcome. For example, Mobleye and Nervana acquired by Intel. In addition, the acquisition of SELLINGS, which has aroused tremendous concern in China, is a deep lesson in science and technology. But in order to be acquired by giants, Chengdu's close relationship with giants and its ability to enhance its ecological strength are crucial. For example, after the establishment of Shenzhen Science and Technology in 2016, it has been developing machine learning solutions based on the technology platform of Sales. The two companies cooperate closely. Of course, start-ups also need their own innovations. At the 2017 Conference on FPGA, the paper was rated as the only best one.

Look at the market and rush in first

Recognizing their own strengths and being able to seize market opportunities and focus on the market are also favorable factors for the success of start-ups, such as Horizon, Yuntianfei and Cambrian start-ups. Horizon was established in 2015, and the first generation of AI chips have been commercialized on a large scale. In the field of intelligent driving, Horizon has established partnerships with Tier1s and OEMs, the four major global automotive markets. In the field of AIoT, Horizon has enabled a number of domestic first-line manufacturing enterprises, modern shopping malls and well-known brand stores.

Yuntian Lifei was also established very early. After two years of its launch in Longgang Public Security Bureau based on IFaaS 1.0 in 2014, Yuntian Lifei took the lead in realizing "millions of people, second-level positioning". After the "deep-eye" dynamic portrait recognition system has become a phenomenal product, Yuntian Lifei streamed a Deep Eye 1000 embedded with independent intellectual property rights in October 2018. Visual AI chip further expands the application scenarios of products to AI edge computing scenarios such as robots, UAVs, smart cities and new retail.

Cambrian is also the first to enter the AI chip market, and with the help of terminal AI smart processors in 2016, it has become a key part of Heise Kirin processor to achieve AI, and of course it should also be used in tens of millions of smart phones. In the Cambrian period of 2018, cloud processing chips were introduced, and the layout of cloud + terminal AI chips was completed.

Technological strength, combined with the first-mover advantage, has led to some AI chip start-ups.

Extending Algorithmic Advantage to Hardware

In the AI era, the importance of hardware-software integration has become more prominent, so in addition to hardware startups, there are also companies that are good at extending algorithms to hardware. Speech is a typical representative. Although Speech was founded in 2007 and focused on the natural voice interaction of intelligent hardware, it did not release the first generation of AI voice chips until early 2019. The logic behind this is that, as an algorithmic specialist, Spitzer used third-party chips to meet the demand before, but with the evolution of voice algorithms and the intensification of market competition, higher requirements have been put forward for AI voice chips. Moreover, self-developed chips can better play the advantages of hardware and software, and more conducive to participating in the market competition.

Touch Infinity also has the advantage of AI algorithm. Since 2016, it has provided practical front-end perception products, and has a better understanding of the pain points of existing chips when landing in security scenarios. In order to further enhance the advantages of front-end perception, Touch Infinity is gradually increasing investment in chips. Together with well-known chip companies in China, we will build artificial intelligence perception chip (SoC) to provide end-to-end comprehensive solutions for the next generation of artificial intelligence perception field. Compared with the existing front-end AI chips, their front-end sensing chips will aim at application landing from the beginning of design, and build chips that really suit the security market with years of front-end experience.

Looking for Unique Products + Market

From image AI chip to voice AI chip, from different angles of hardware to algorithm, the startups mentioned above are mainly in terminal AI chip market. Because the ecological advantages of the giants in the cloud-based AI chip market are hard to break, most start-ups still have to abandon this market even though the cloud-based AI chip market is more profitable and the market demand is more stable.

But there are also exceptions, such as Tian Number Intelligence Core, which was founded in 2015. They neither focus on AI algorithms nor on hardware companies alone, but are positioned as a system-level company that combines hardware and software. Starting from improving software computing power, Smart Core regards software as the key to carrying its ecosystem, and quickly introduces specific landing application solutions in industry scenarios, gains an entrance for chip products, and timely launches AI chips. By means of software and hardware, Smart Core not only improves the average value of computing power, but also increases the peak value of computing power.

At the same time through transparent migration

Among the technological giants who have chosen to develop AI chips independently, Google has taken a step faster. In 2011, Google began to seriously consider the use of in-depth learning networks, but the high demand for these networks makes their computing resources tense. In May 2016, Google unveiled its self-developed TPU for the first time at the I/O conference and said that the chip had been in use in Google's Data Center for one year. TPU 3.0 released in 2018 can achieve 100 PFLOPS performance, energy efficiency can reach several dozen or even hundreds times of GPU.

After Google, foreign Facebook, Amazon, Baidu and Ali have begun to research AI chips by themselves, which has become a highlight of this upsurge of AI. The giants hope to meet their personalized needs at a lower cost by using self-developed AI chips, and ultimately gain greater advantages.