"If the artificial intelligence chip developed with memristor technology is applied to the mobile phone, the power consumption of the chip will be greatly reduced, and the mobile phone can be used for two days with a single charge." On October 27, in an interview with reporters, Tsinghua University Qian He, a professor at the Institute of Microelectronics, explains the changes that memristors will bring to the lives of ordinary people.

At present, artificial intelligence is developing rapidly. However, in Qian He's view, if you want to develop more powerful artificial intelligence, that is, human-like robots, showing the same skill and flexibility as human beings, the role of memristor can not be ignored. The memristor and the human brain have similar synaptic functions, which is equivalent to an "electronic synapse". The built-in smart chip has the ability to learn online and can handle tasks that the machine system could not perform before.

"The memristor is a new type of electronic component whose characteristics are similar to those of the basic unit of information processing in the human brain." Qian He explained that the most unique function of the synapse is that it can be stored or calculated. . The integrated computer architecture is a disruptive technology, and the latest research finds that the memristor is particularly suitable for the integration of deposits and demerits, which is also the focus of research and development of relevant scholars in the world today.

As one of the key projects of the national key research and development program “Nano Science and Technology”, Qian He’s “Basic Research on 3D Integration of New Nano Memory” aims to provide independent intellectual property rights and prototype technology for China's memory industry, supporting and leading China's memory industry. Leapfrog development.

Qian He said that in addition to storage, memristors have shown more important potential in the field of neuromorphic computing chips. The neuromorphic calculation is a new calculation method that greatly improves the computing power and energy efficiency by imitating the structure of the human brain.

The most essential difference between neural morphology calculation and traditional calculation method is that it combines the units responsible for data storage and processing, thus eliminating the large amount of data caused by frequent movement of data between the memory and the central processing unit in the traditional calculation method. Energy costs have bypassed the “storage wall”, a bottleneck that restricts the development of traditional computing methods, and greatly increases the parallelism of data transmission and processing.

"The resistance of the memristor itself can be used to store data; the corresponding current is also output under the applied voltage to complete the multiplication calculation function, and the output currents of the plurality of memristors can be combined to achieve the function of addition calculation. Through the combination of multiplication and addition, the memristor can perform most of the calculation tasks in a very short time." Qian He said that especially the resistance of the memristor can be fast and reversible under certain applied voltage conditions. The control and adjustment makes the neuromorphic computing chip integrated by the memristor not only able to perform computational functions efficiently, but also reprogrammable. These characteristics bring unparalleled advantages to the application of memristors in the field of neuromorphic computing.

"It is foreseeable that once the memristor-based neuromorphic computing chip technology matures, making a 'super artificial brain' that resembles or even surpasses human brain intelligence and energy efficiency will become a reality." Qian He said.

According to reports, the research team completed the world's first integration of thousands of bidirectional continuous resistance memristor units by exploring the working mechanism of memristors, screening materials, optimizing structures, carefully designing chip circuits and test systems. A face recognition system that has real-time response to input and can learn online like a human brain, and the system consumes less than one-thousandth of the energy of the Intel Xeon Phi coprocessor. Relevant results have been published in the May issue of Nature Newsletter.

“We integrated a computing chip with 1024 memristor units, which perfectly realized the emotion recognition function of face images and obtained more than 90% recognition accuracy on the more difficult test pictures with 20% noise points. Rate." Qian He said.