The world's first memristor chip, what breakthrough did the Tsinghua team make?

The wave of development of artificial intelligence technology has greatly changed human production and lifestyle, and has had a profound impact on all areas of current human civilization, and will continue to have a profound impact.

However, artificial intelligence technology relies on the support of large computing power. With the explosive development of technology, its demand for large computing power is also rising. However, the existing shortage of computing power and the huge demand for computing power have formed an increasingly prominent contradiction. As the material carrier of computing power, chips are facing challenges that need to be overcome urgently. How to speed up the development of high-computing power and high-efficiency chips, solve the huge computing power gap, and achieve a substantial increase in computing power is an urgent problem that needs to be solved by current hardware technology, and it is also the core problem that is stuck in the "neck" of the development speed of artificial intelligence technology.

Recently, the team led by Professor Wu Huaqiang and Associate Professor Gao Bin from the School of Integrated Circuits at Tsinghua University developed the world's first fully integrated memristor memory-computing integrated chip that supports efficient on-chip learning (machine learning can be done directly on the hardware end) based on the memory-computing integrated computing paradigm. This has made a major breakthrough in the field of memristor memory-computing integrated chips that support on-chip learning, and is expected to promote the development of artificial intelligence, autonomous driving, wearable devices and other fields. The relevant results were published online in the latest issue of Science under the title "Edge Learning Using a Fully Integrated Neuro-Inspired Memristor Chip".

Image source: unsplash

What is a memristor chip? It is said that memristor is the fourth basic circuit element after resistor, capacitor and inductor.

Traditional chips before the invention of memristor chips were based on the von Neumann model, which separated memory and processor and connected them through a data bus, requiring data to be moved back and forth between the processor and memory. The high energy consumption and high latency, low privacy and security, low adaptability and low robustness brought about by this separation of storage and computing have become a major challenge restricting the improvement of computing power.

In order to solve these problems and challenges, a new computing paradigm has been proposed, namely the storage-computing integrated computing paradigm. The storage-computing integrated computing paradigm refers to integrating the memory and processor on a single chip and using the physical properties of the memory itself to perform computing. This solves the shortcomings of the traditional computing architecture paradigm.

Image source: unsplash

The key to the integrated computing paradigm is that the memory itself has computing capabilities. To achieve this, a new type of memory device was invented, the memristor. A memristor is a resistor that has memory functions like neurons in the human brain. After power is cut off, it can still "remember" the charge that passed through it before. It is a major breakthrough in the field of electronics and has shown great potential in data storage, computing, encryption and communication.

Since 2012, the team of Qian He, Wu Huaqiang, and Gao Bin from Tsinghua University started from the research and development of memristor devices and prototype chips, and gradually developed to system integration and computing theory. They developed the world's first fully integrated memristor storage and computing chip that supports efficient on-chip learning. All learning-related calculations are completed on this chip.

Image source: unsplash

The hardware test results show that the chip contains all the necessary circuit modules to support complete on-chip learning. In multiple on-chip learning tasks involving image classification, speech recognition, and control tasks, the chip's energy consumption is only 3% of that of an ASIC system under advanced technology. At the same time, it is expected to achieve a 75-fold increase in energy efficiency and can effectively protect user privacy and data. It demonstrates high adaptability, high energy efficiency, high versatility, and high accuracy, and has great application potential to meet the high computing power requirements of the artificial intelligence era.

References

[1]https://www.science.org/doi/full/10.1126/science.ade3483

Planning and production

This article is a work of Science Popularization China-Starry Sky Project

Produced by: Science Popularization Department of China Association for Science and Technology

Producer｜China Science and Technology Press Co., Ltd., Beijing Zhongke Xinghe Culture Media Co., Ltd.

Author: Zeng Xinyue, popular science creator

Review丨Huang Yongguang Associate Researcher of Optoelectronic Chips, Institute of Semiconductors, Chinese Academy of Sciences

Editor: Zhong Yanping and Qi Yuan (Intern)