Tsinghua has developed a chip that can support neural networks. This is a major breakthrough. With the turbulent AI hardware boom, China's semiconductor industry has also ushered in a rare opportunity. Computer chips are the key to AI success and can support neural network chips. It is still only a microcosm of China's scientific and technological circles. It will take some time for China to realize a true technological power.
According to foreign media reports on January 24, the research team of Tsinghua University has made a major breakthrough and developed a chip that supports neural networks, which can be applied to small devices using batteries.
In an office at Tsinghua University in Beijing, the "Thinker" chip is processing camera data to find faces stored in the database. After a few seconds, the same chip is processing Chinese voice commands. "Thinkers" can support neural networks, and they are unique in that they are powered by a low-energy drive - eight AA batteries can supply the amount of power they need for a year. The Thinker can dynamically adjust its calculation and memory requirements to meet the software needs. This is important because many existing AI applications—either to identify objects in a picture or to understand human speech—require a combination of multiple neural networks and multi-layer neural networks. In December 2017, the paper introducing the "Thinker" design was published in the IEEE Journal of Solid State Circuits, the top journal of computer hardware design. This is an amazing achievement for the Chinese research community. Moreover, although AI software can be run using existing graphics processing chips or FPGAs (a blank chip that can be dynamically reconfigured), those designs are expensive and cannot be used in small devices that use batteries. This is why Yin Shouyi’s team wants to develop “thinkersâ€. Thinkers can be embedded in a variety of devices, such as smartphones, watches, home robots, or devices that are stationed in remote locations. Yin Shouyi’s team plans to release the first product to match the “Thinker†in March this year.
This chip is just a microcosm of the Chinese scientific and technological community. In the optimisation of AI hardware, China's semiconductor industry has also ushered in a rare opportunity to establish its position. Computer chips are the key to AI's success, so China needs to develop its own hardware industry to become a true technology power. Yin Shouyi, deputy director of the Department of Microelectronics at Tsinghua University and the lead author of the "Thinkers" paper, said: "China has kept up with this tide at the fastest rate compared with the changes in information technology before the response."
Even though China has become a manufacturing center for solar panels and smartphones, its semiconductor industry still lags behind the US. From January to September 2017, China invested US$182.8 billion in imported integrated circuits – an increase of 13.5% year-on-year. In the "Three-Year Development Plan for Artificial Intelligence" released by the Ministry of Industry and Information Technology in December 2017, the Chinese government has set a goal of capable of mass production of neural network processing chips in 2020.
In addition, there are other similar programs in China. At the end of January, a research team at the Institute of Computer Technology of the Chinese Academy of Sciences will work with local semiconductor manufacturers to produce a small number of chips for robotics. The chip is called "Dadu" and has two cores - one for running the neural network and the other for controlling the action. The neural kernel runs the visual algorithm, but does not affect the action kernel to plan the best route or the best take action. Han Yin, director of the Network Computing Laboratory of the Chinese Academy of Sciences and head of the robotic chip project, envisioned a range of applications, including coffee robots and drones with gesture control. The advantage of developing such a system in China is that the user base is wide, which makes it faster to upgrade the chip based on the user experience.
China's IC industry is also expanding rapidly, accounting for 58% of the world's IC market growth from 2006 to 2016. However, according to PricewaterhouseCoopers, in 2016, China's share of the world's semiconductor manufacturing capacity was only 14.2%. In 2015, China's central government issued the manufacturing policy "Made in China 2015", chip design and manufacturing is one of the key areas of major breakthroughs required by the government.
However, Chinese chip companies have found that their environment is completely different from the environment that fostered Intel and NVIDIA. SmarCo, based in Beijing, designed AI chips for data centers that process video clips. Its president, Fan Dongrui, said that cloud computing dominates most markets, leaving little room for finished hardware. Despite this, more and more Chinese AI companies are developing their own hardware. Ma Fengxiang, director of application-specific integrated circuit design at Horizon Robotics, said that "only companies that make chips will be less and less in the future." In December 2017, the company released two computer vision chips that could allow cars to identify pedestrians or Help the mall find customer traffic patterns. Ma Fengxiang said that the horizon robot is not a chip company, but will design chips for internal products to improve performance and reduce production costs.
Now, there are still many problems that need to be solved by Chinese chip research experts: how to realize the commercialization of chip design, how to expand the scale and how to lead a computer world transformed by artificial intelligence. But what researchers are most lacking is ambition. Han Yinhe said: "As a chip researcher, we all have dreams. Regardless of the progress of the research, we will wait and see."
The problem
Throughput is an important indicator of Ethernet testing. Many engineers believe that the Ethernet switching throughput should be its line rate, that is, no packet loss can occur at 100% traffic, and that it is illegal for the Ethernet frame interval IFG to be less than 96 bits. However, in the Ethernet switching throughput and packet loss rate test, a small amount of packet loss occurs in the long-term error test under the line speed condition. The reason is the cross-clock domain architecture of the Ethernet.
With the rapid development and application of industrial Ethernet technology, a large number of network problems have appeared along with it. According to statistics provided by Siemens, the network communication failure rate accounts for more than 70%, and the network equipment failure rate is less than 30%. After a network failure causes the system to shut down, the time required for fault diagnosis and location accounts for more than 80% of the total system downtime, while maintenance measures take up less than 20%. Therefore, real-time monitoring and analysis of network traffic is a major problem in the development and application of industrial Ethernet. Real-time monitoring and analysis of industrial Ethernet network traffic and timely detection and location of network problems play a vital role in improving the stable operation of the entire system. .
100M Switch,Fast Ethernet Switch,Fe Port,100M Ethernet Switch
Shenzhen GL-COM Technology CO.,LTD. , https://www.szglcom.com