Economic Information Daily
There is a saying that the present artificial intelligence is particularly suitable: 'We tend to overestimate the present and underestimate the future.' If we look forward to it for a period of more than 60 years now, it really swept through human society and it has just begun.
From such a long historical point of view, from the perspective of the future of general artificial intelligence, these pioneers who occupy the stage center in this era all hold the most primitive tools such as 'deep learning' and 'neural network' Wading in the wilderness before walking on the road to find the basic tools of artificial intelligence, Chen Tianshi no other than what the prerequisites.
Fortunately, from the stone tools he created, people gradually discovered and kept the fire.
Cambrian, this backed by the CAS Institute of Computing startups just debut in the market, entered the 'Unicorn' field.
Polished 'stone'
AI is a revolution. Like the Internet, it swept through everything, from nothing to nothing (everywhere). Ordinary people do not even need to go forward, it will take the initiative to come and hug you, and even coerce you In the words of Luo Zhenyu: 'You do not have to rush because you will eventually arrive.'
In this revolution of redefinition of infrastructure, Chen Tianshi is undoubtedly an early departure. 16-year-old admitted to the University of Science and Technology of China juvenile class; 2014 papers in the United States held the best international academic conference paper award; founded Company Cambrian Technology, A round valuation of more than 1 billion US dollars, becoming the world's first artificial intelligence chip 'Unicorn', the world's largest AI chip startups.
Based on the underlying algorithm of deep learning, great progresses have been made in such fields as computer vision, speech recognition, natural language processing, audio recognition and bioinformatics.However, the traditional processors such as CPU are the mainstream AI computing platforms at present, However, the basic framework of the structure is not designed for AI, the efficiency is subject to many limitations, including the GPU also has a large power consumption and other defects.
If you want to use a general-purpose processor to build a brain-sized synaptic neural network, you may need to build a power station to power it.Alfa dog just appeared, the next game with 1000 CPUs and 200 GPUs, the electricity bill per minute Up to 300 dollars, while the network size is only one-thousandth of the human brain.
Chen Tianshi believes that the ideal AI chip should be a new type of processor that has a wide range of applications (including voice, semantic, image, video, natural language multi-modal processing capabilities), while far exceeding the CPU and GPU Efficiency.To achieve this goal, you must have a new AI instruction set, the use of instructions for flexible processing, it is possible in the AI chip to a variety of algorithms applications can support faster and better.
For a simple example, the motor of a motorcycle (PC, cell phone, etc.) is mounted on a car (deep learning platform), and the car naturally does not work well; now it is time to design a car specifically for the engine Device).
DianNao's depth of learning processor architecture, the average simulation performance than the mainstream CPU core of 100 times, but the area and power consumption of only 1/10, the performance up to three orders of magnitude.In 2014, he and his brother Chen Yun and This essay in collaboration with Prof. Olivier Temam of Inria is the best paper for a leading international conference.
Between 2014 and 2016, they and their international academic collaborators almost swept the academic framework of the processor architecture. At the top academic conferences, they taught Diannao series of foreigners to speak Chinese: Diannao, DaDiannao, PuDiannao, Computer), ShiDiannao (Computer), Cambricon (Cambrian Instruction Set) and a series of innovations that aim to improve the efficiency of current deep learning by hundreds of times for different purposes.
According to some academicians, the groundbreaking progress made by the 'Cambrian' in the deep learning processor instruction set has provided technical support for our country to occupy a leading position in the smart industry ecology. Since 2014, the CAS Institute of Computing and Inria Joint Team proposed Depth Learning Processor architecture DianNao, Deep Learning Processors has become one of the ISCA's most focused research topics at ISCA International Conferences. Nearly one-sixth of ISCA's 2016 papers cite 'Cambrian' work for depth Learn processor exploration.
'At present, software algorithms like Alpha Dogs are executed on the GPU, and deep learning (neural networks) will be significantly faster with the Cambrian processor in the future,' said Chen Tianshi.
He used the 'Swiss army knife and chopper' to describe the relationship between a generic processor and a deep learning processor: Although the Swiss Army knife is more versatile, it is more suitable for cooking when cooking, and in terms of intelligent handling, 'Cambrian' is Put a more suitable chopper.
The Cambrian instruction set deals directly with large-scale processing of neurons and synapses. One instruction completes the processing of a group of neurons and provides a series of specialized support for the transmission of neurons and synapse data on the chip Simulations have shown that deep learning processors using the Cambrian instruction set have two orders of magnitude performance improvements over the x86 instruction set central processing unit.
Now, by simulating neurons and synapses in computers, the Cambrian AI chip intelligently processes the information, handling 16 billion neurons and more than 20,000 per second with a specially designed memory structure and instruction set 100 million synapses, power consumption is only one-tenth of the original, even in the future there is even hope that the system similar to the Alpha dog into the phone.
Buried heavy industry
Chip is the information industry's 'Daojian even', but also 'heavy industry', its investment required is huge, far more than other areas of AI.
In August 2017, Cambrian Technology won A round of financing worth US $ 100 million. This round of investment is led by SDIC Venture Capital, Alibaba Ventures, Lenovo Venture Capital, NSG Investment, Zhongke Turing, Yuanhe Origin (Angels round collar investor), Chung Hua investment (Angel wheel investors) joint investment.
This backed by the CAS Institute of Computing startups just unveiled, to enter the market, entered the 'unicorn' field.
Huawei's flagship mobile phone Mate10 uses the Cambrian 1A processor, which is the first step in the industrialization of the Cambrian AI industry. This 2016 release of the world's first commercial depth learning processor IP, selected the third world Internet Conference selected the top 15 "world's leading Internet technology achievements."
It not only has completely independent intellectual property rights, but also has outstanding versatility and performance ratio in key artificial intelligence tasks such as computer vision, speech recognition and natural language processing, achieving the performance of more than 25 times of traditional quad-core general purpose CPU and more than 50 times Energy efficiency.
'Our product, the Cambrian 1A processor, can compete with Apple products, and as a result we can easily win the fight. We do not rely on brute force to pile up calculators, but rather rely on the most advanced, Sparse technology, deftly deflected, so Huawei Mate10 intelligent processing speed beat the powerful Apple iphoneX. "Chen Tian Shi said.
Under the guidance of the ultimate direction of universalizing AI, he set his own small goal. 'We are now on the same starting line with our foreign counterparts and hope that Cambrian will occupy China's high-performance smart chip after 3 years Market share of 30%; and three years later, there will be one billion smart devices integrated with the Cambrian processor in the world. "
He gave the path to achieve.Carbon Marcus chips on the commercialization of the market: 'First, the terminal, the second is the cloud end products are smart phones, smart glasses, drones, autonomous vehicles, etc., need to identify the chip Images, videos and texts, while in the cloud, such well-known cloud clients as UCITS (52.750, 0.34, 0.65%) and Zhongke Dawning (39.400, 1.59, 4.21%) are already customers of the Cambrian era. '
For this small goal, he said he was 'working at rest except for rest' and ran furiously in the top of the smartest application area, using face recognition and semantic understanding to systematically integrate Autonomous driving, smart phones, etc. The 'iceberg' is invisible to the technology.
The chip is the 'foundation' of the technology layer. 'The chip is something like building a house or building a road. If there's no way for the house to go forward, I hope we can continue to take on the role of infrastructure providers in the smart era Character. "Chen Tianshi said.
Technology must bring the gulf between people, and lowering the threshold of technology and creating as fair a fair as possible are the only viable ways to eliminate this glitch.
'From an individual point of view, we are mainly engaged in the development of artificial intelligence acceleration chips that allow developers to get cheaper and more efficient chips and support the development of higher-level applications.' 'Chen Tian-shih said that only if the entire computing system is cheaper would it be possible to promote artificial Therefore, at the bottom, reducing power consumption and improving efficiency are the keys.
As for the upper application, the wisdom of the masses is endless.Chen Tianshi is optimistic, he holds a happy attitude to support the development of artificial intelligence, I believe in the basic layer, grow more applications.
Keep 'fire'
In the end of 2017, a manuscript entitled "Chip-and-Chip Times" caught fire in the AI world. According to CB Insights, the total investment in emerging chip startups has risen from 800 million U.S. dollars in 2015 to 16 in 2017 In the second half of 2017, more than 30 AI chips were waiting in line at TSMC's production line for filming.
When the potential also can not reach the end, not only depends on the speed of running, but also depends on where the track.
Not only started the company running, in fact, the giants awakened faster than anyone else.On the previous two days, chip giant NVIDIA held a conference at the Consumer Electronics Show, announced a series of unmanned vehicle machine learning chip Xavier related In view of the performance in the field of AI chips, its stock rose nearly 100% in 2017, making it the 'Best Wealth Product of the Year'.
Not to mention Intel, Google, Facebook, Microsoft and other giants are doing all this.Consulting firm Tractica forecast data show that by 2025, artificial intelligence-related deep learning chipset market revenue, from the previous 500 million US dollars Soared to a scale of 12.2 billion U.S. dollars, a compound annual growth rate of over 40%.
Life and death fight, 100 billion market victory in the battle.
'The success or failure of the chip, in addition to its own efficiency, the ecology is a crucial link.No matching applications and software, it is difficult to succeed in the market. "Chen TianShi said that as long as the domestic AI command to live in, China dominates the world AI industry opportunities may come.
At the first press conference of Cambrian Technology, he took a long time to introduce his partners, saying that in the past our domestic enterprises started late and was always the follower of the international ecology and we have integrated into this ecology, Enjoy this bonus, but we do not have much power to change this ecology.However, in the smart age, everyone is on the same starting line, dare to be the best in the world, dare to come forward and jointly build a new smart ecology. Wu Ji is not enough. We all need to work hard together on the hardware instruction set and the software development platform, work hard together in the cloud and work together in business collaboration. '
On October 16, 2017, Chen Tianshi made a circle of friends and said, "He is a matter of life, but he is not a teacher." The phrase from Gong Zizhen was quoted by many educators .
Afterwards, he remembered that this was probably because of his stubbornness. 18 years ago, Chen Tianshi was admitted to China University of Science and Technology, and later under the direction of Academician Chen Guoliang and Professor Yao Xin, Ph.D. in artificial intelligence. Things, the bottom of the ecological cornerstone is the chip to do well. 'Technology must be the best.'
In the past, it is now, in the future, no matter where the company goes, technology will always be the number one priority.