Scientists develop more efficient memristor chip

Chinese scientists have developed a fully integrated memristor chip with improved learning ability and low energy cost, according to a study recently published in the journal Science.

With artificial intelligence technology profoundly changing the methods of production and life, learning has become highly important for intelligent devices to adapt to different application scenarios.

However, current technologies for training neural networks require moving extensive data between the processor chip and off-chip main memory, which incurs massive energy consumption and hinders the learning process.

Based on 11 years of research, scientists from Tsinghua University developed a full-system-integrated chip consisting of multiple memristor arrays and all the necessary peripheral circuits to support complete on-chip learning.

Memristor is an electrical component that regulates the flow of an electrical current.

"The chip integrates complete circuit modules to support autonomous learning, and it has successfully demonstrated various learning tasks including motion control, image classification and speech recognition," said Yao Peng, the co-first author of the study, from the School of Integrated Circuits, Tsinghua University.

According to the study, the chip can achieve autonomous learning with only about three percent of the energy consumption of the conventional application specific integrated circuits when running the same task.

The chip showed high energy efficiency and high accuracy in versatile AI tasks during the experiments, which can effectively strengthen the learning adaptability of intelligent devices in practical application scenarios, Yao said.

"This chip provides an innovative development path for AI hardware breakthroughs," said Gao Bin, a professor at Tsinghua University.

This study is an important step toward future chips with high energy efficiency and learning capabilities, said Wu Huaqiang, dean of the university's School of Integrated Circuits.

"We hope our findings will accelerate the development of future smart edge devices that can adapt to different application scenarios and owners," Wu added.

Xinhua