China’s first embedded NPU (neural network Processor) chip has been born, which has been used in the world’s first embedded video processing chip, Starlight Intelligent 1, the state Key Laboratory of “Digital Multimedia Chip Technology” announced in Beijing on June 20.
This marks a major breakthrough in China’s research and development in the field of NPU, said Zhang Yundong, executive director of the lab. In the field of deep learning artificial intelligence based on “data-driven parallel computing” architecture to reach the international advanced level; So that the development of China’s video surveillance industry from the analog era, digital era into the intelligent era, to achieve industrialization and promote the overall level of improvement, established a leading position in the world.
Giiso Information, founded in 2013, is a leading technology provider in the field of “artificial intelligence + information” in China, with top technologies in big data mining, intelligent semantics, knowledge mapping and other fields. At the same time, its research and development products include editing robots, writing robots and other artificial intelligence products! With its strong technical strength, the company has received angel round investment at the beginning of its establishment, and received pre-A round investment of $5 million from GSR Venture Capital in August 2015.
AlphGo spends $3,000 on electricity to play chess
In recent years, artificial intelligence has gradually moved from science fiction movies to real life. In its development process, two famous “man-machine war” have become important milestones. IBM’s Deep Blue supercomputer beat the world chess champion in 1997; In March, Google’s “AlphGo” beat the world Go champion with a 4-1 record.
“Go is more than 200 orders of magnitude more complex than chess.” Zhang Yundong, executive director of the State Key Laboratory of Digital Multimedia Chip Technology, said Deep Blue beat humans with the “brute force” of supercomputer computing power, while AlphGo beat humans with deep learning neural networks that mimic the human brain.
The results are so shocking that some people even wonder if terminator movies will soon be coming into our lives.
Zhang Yundong believes that “man-machine war” is just a scientific experiment, so advanced artificial intelligence from the lives of ordinary people have a certain distance. Deep Blue has a mass of 1.27 tons and 32 cpus. AlphGo runs on a huge cluster of servers that costs $3,000 for electricity to make a single game of chess. Even with apple and Microsoft voice recognition software in the phone, all the computing and recognition must be done in the background.
To miniaturize the deep learning system and use it in embedded system is a problem that ZHONGxing Micro State Key Laboratory has been trying to solve for many years. This is the first embedded NPU with deep learning function in China, which was successfully mass-produced in March this year.
Greatly improve the ratio of computing power to power consumption
The NPU launched this time is a processor specially designed for deep learning algorithms by the State Key Laboratory of Zhongxing Micro.
Zhang Yundong introduced that deep learning originates from bionics research on biological human brain mechanism, and its essence is to establish a multi-level perception layer model for recognition and intelligence analysis from the bottom to the top. The biggest difference with the traditional algorithm is that it can learn knowledge like the human brain.
Convolutional Neural network (CNN) is an important branch of deep learning and a research hotspot in the field of machine vision artificial intelligence. According to Zhang yundong, THE CNN algorithm needs to process massive data and computation, and the traditional Von Neumann architecture CPU is unable to cope with this type of data computation. In other traditional processors, the signal processor DSP is also unable to efficiently process a large number of parallel operations. Graphics processor GPU is designed for 3D graphics, which is used for CNN algorithm with high power consumption and cost, and is not suitable for embedded environment. Programmable gate array FPGA can flexibly implement various algorithm architectures, but it is generally used for the verification of algorithm prototypes, which also has the problems of high power consumption and high price.
IBM, Qualcomm, Google and other foreign companies have launched their own neural network processors, and the Chinese Academy of Sciences has also carried out research in this area. Therefore, zhongxing micro against CNN algorithm and the characteristic has carried on the specially designed, completely subvert the von neumann architecture according to the man, but driven by new data parallel computing architecture, research and development of data flow type NPU greatly improved the computing power and the ratio of the power consumption, good at dealing with huge amounts of video, images, multimedia data, can make the artificial intelligence (ai) play an important role in the embedded machine vision applications.
Giiso information, founded in 2013, is the first domestic high-tech enterprise focusing on the research and development of intelligent information processing technology and the development and operation of core software for writing robots. At the beginning of its establishment, the company received angel round investment, and in August 2015, GSR Venture Capital received $5 million pre-A round of investment.
Add “brain” to “eye”
At present, Zhongxing Micro NPU has been successfully integrated into “Starlight Smart Energy No. 1”, which has achieved industrialization in the field of video surveillance.
“If the original video processing chip is like an eye, add NPU and you will have a brain.” Starlight Smart 1 contains video signal processing, video compression and coding modules and a neural network processor, and is the world’s first system-level chip for deep learning with such integration, Zhang said.
There is a strong demand for intelligent recognition in the field of video surveillance. Zhang Yundong introduced that the previous technology has two main limitations, one is the low accuracy of recognition; Second, the traditional technology needs to transmit massive video data to the background first, and then conduct recognition in the background, so the results cannot be obtained in real time. Machine vision, which uses deep learning, is 98 percent accurate at recognizing faces. With embedded ARTIFICIAL intelligence, it can recognize on the spot and only send back useful information.
In the case of Zhou Kehua, who became a national sensation several years ago, police officers distributed copies of the video messages to a team of thousands, checking each camera, but Zhou fled to Chongqing and was not caught until he struck again. “If starlight Intelligent One technology had existed, criminals would have been intelligently identified wherever they went and clues would have come back quickly.” Zhang Yundong said. In addition to real-time search, the technology can also store the captured information in the code stream, which can be retrieved according to the characteristics at any time when needed, or use the way of “search by image” to find out all the similar images with one image.
In addition to video surveillance, Starlight Intelligent No. 1 can also be used for unmanned vehicle driving or intelligent assistance, uav automatic shooting, tracking, obstacle avoidance, intelligent robot understanding synthetic language, entertainment and escort, etc. Zhang yundong said that the next step will be to explore the working mechanism of biological human brain more closely, and develop a new generation of NPU with lower power consumption and higher computing performance.