At the same time, a number of outstanding smart chip companies have emerged in the past two years in China, such as the Cambrian, Horizon, Yunzhisheng, Shenjian Technology, Zhongtianwei, etc. There are also some large companies such as Huawei, Ali, etc. In one area, this puts China's smart chip field in the international first camp. But looking around the world, Nvidia, Google, Intel, ARM, Qualcomm, Xilinx and other traditional big-name companies are also in the smart chip field.
Since the beginning of this year, various companies have successively launched their own smart chip products, which indicates that the research and development period of smart chips has passed and is gradually entering the industrial promotion period. Although each has not yet mass-produced, there is a war for market competition. Or quietly started. Whether it can win the market may not only determine the survival of an enterprise, but also determine whether China's chip industry can achieve catch-up and breakthrough.
The rise of smart chips
With the re-emergence of artificial intelligence, smart chips have been greatly developed in recent years.
What is a smart chip? Just as "smart" is not well defined, it is very difficult to give a precise definition of a "smart chip." CCAI2018 Sub-forum "Smart Chip" Chairman, Han Yinhe, a researcher at the Institute of Computing, Chinese Academy of Sciences, believes: Above all, chips that can provide special intelligence capabilities can be called smart chips; and what we currently call smart chips are chips that are tailored to artificial intelligence algorithms for devices, circuits, or architectures, especially for deep learning. Smart chip.
The GPU was originally a microprocessor chip for dedicated image computing on personal computers, workstations, game consoles, etc. Later researchers found that their massive data parallel computing capabilities are highly matched to the deep learning application characteristics. In 2011, Wu Enda took the lead. The GPU was applied to Google’s brain and achieved amazing results. The results showed that 12 NVIDIA GPUs can provide the equivalent of 2,000 CPUs for deep learning. Later, researchers from New York University, the University of Toronto, and Swiss artificial intelligence labs were Accelerating its deep neural network on the GPU. As a leader in the GPU industry, Nvidia has quickly become a giant in artificial intelligence, achieving a six-fold increase in stocks a year.
On the other hand, with the rise of deep learning, many scholars have thought of developing deep learning accelerators, that is, accelerating the computation of neural networks through hardware implementation. Since 2009, Y. LeCun, O. Temam et al. in the convolutional neural network. After some preliminary work on accelerator design, from 2014 to 2016, the research team of the Chinese Academy of Sciences Chen Yunbiao successively published several papers on deep learning accelerators in ASPLOS, MICRO, and ISCA, the top conferences in the field of computer architecture. The academic community’s enthusiasm for deep convolutional neural network accelerated chip research. Based on these studies, Chen Yunqi and his DianNao chip have achieved 100x performance acceleration. Then in March 2016, Beijing Zhongke Cambrian Technology Co., Ltd. registered. Founded, based on the technical framework of the DianNao project, the Cambrian launched the "Cambrian 1" chip, the Cambrian 1A processor and other products. The latter is the earliest mass-produced AI chip. In 2017, Huawei's first artificial intelligence chip, the Unicorn 970, integrated a neural network processor. 970 has been equipped with a unicorn on Huawei Mate 10, P20, V10 glory of the three series of mobile phones, cumulative shipments have reached tens of millions.
In the same period, Google also saw the great potential of deep learning in practical applications. Unlike DianNao's architecture of multiply-sum tree architecture, Google's tensor processor TPU for data centers uses the organization of pulse arrays. The systolic array was proposed by Professor Kong Xiangzhong of Harvard University in the 1970s. In May 2016, Google announced the first-generation TPU for the first time at the I/O Conference, and introduced that TPU is one of the “secret weapons” that AlphaGo can defeat Li Shishi. At the beginning of 2018, Google announced the opening of its TPU cloud service platform, which was priced at US$6.5/hour. However, based on the needs of its business model, Google currently researches and develops its TPU only for its own internal use and does not intend to sell chips.
Of course, there are many kinds of smart chips. The chip's application scenarios are different, and its design is also different. For example, Google’s TPU is designed based on their cloud computing application scenarios. Its power consumption is relatively large, but it is more important for it. The performance should be high enough. For another example, Huawei's Unicorn 970 requires low power consumption and moderate performance because it is embedded in mobile phones; while the horizon-oriented chips for unmanned design require special treatment for vision.
Since 2014, the research on artificial intelligence chips has been for four years. In this field, a number of companies have emerged in our country, such as the Cambrian, Horizon, Yunzhisheng, Shenjian Technology, Zhongtianwei, etc.; These companies have also launched smart chip products that are adaptive to the scene. In addition to the Cambrian period described above, the “Sunburst 1.0” and “Jourge 1.0” that Horizon announced at the end of 2017 are mainly for smart cameras and smart driving. Therefore, in this wave of smart chips, our country does not at least appear to be falling behind.
The next war
According to Han Yin and researchers, with the advent of multiple smart chip products this year, smart chips will gradually go through the research and development period, but “release” does not mean “application.” Now except for the Cambrian, Huawei’s chips, and other domestic The chips of the manufacturers have not yet been used on a large scale. From an international point of view, the products of Intel and other companies have not yet been mass-manufactured. At present, only Nvidia's GPUs occupy the market. Therefore, the current state seems to be that each company is The prototype chip was launched sooner or later, but it has not yet been played; in other words, everyone is already on the runway, but they have not yet started to run.
But, perhaps this is the eve of the storm. The next battle will be a battle for market competition.
Looking now, China's smart chips are not easy and there are no challenges. Older companies such as Google, Intel, and ARM have developed their own smart chips in different smart chip applications. They will all be competitors of our chip companies. For example, in cloud computing On the other hand, Google’s TPU has been researched for many years. Although it has not been sold yet, once it is decided to sell, it will certainly have a huge impact on China’s cloud computing market. In the high-performance market, Intel’s history has never been ignored. Any potential competitor, this field has many challenges in history, but it was defeated by Intel. Then there is ARM. Even though Huawei has already launched its own product in the field of mobile phone chips, now ARM is also developing their neural network. Accelerator, if it integrates with its ARM core in the future, its influence will not be underestimated.
Therefore, although we are not lagging behind in the research and development phase of smart chips, if we still want to maintain this advantage in applications, we will face many challenges. From research and development to application, it is a long road, and many problems remain to be solved. .
What are the latest developments in smart chip R&D and market in China? What are the smart chips for the smart market? How do smart chips serve the needs of mobile phones, security, home appliances, and automobiles? How to use China's application advantages to establish ecological advantages? For countries in the smart chip industry What are the policy recommendations in the development?
“If you don’t want to be a global player, you will not be able to seek a domain.” The smart chip is the core engine of the smart era. In response to these hot issues, the China Artificial Intelligence Society will hold the China Artificial Intelligence Conference (CCAI2018) in Shenzhen from July 28th to 29th. The “Intelligent Chips” sub-forum was set up, inviting leading companies such as the Cambrian, Yun Zhisheng, Shen Jian and other leading domestic smart chip makers, as well as well-known scholars in the international chip field, as well as whole machine manufacturers and application vendors to discuss these issues. Can provide some new ideas for the industry.
Conference registration address: https://www.bagevent.com/event/1449943?bag_track=wangyizhineng