Chinese Scientists Develop Super-efficient All-analog Photoelectronic Chip
Researchers from Tsinghua University have developed an allanalog photoelectronic chip that can process computer vision tasks with greater speed and more energy efficiency than existing chips.
The research team's findings, which provide an alternative to existing technologies based around analogue-to-digital conversion, have been published in the journal Nature.
Analog and digital signals are two types of signals. Analog signals vary continuously, as with the rays of light forming an image, while digital signals are non-continuous, as with binary numbers.
In vision-based computing tasks like image recognition and object detection, signals from the environment are analog, and they need to be converted into digital signals for processing by AI neural networks — systems trained to recognize patterns and relationships in a data set. However, the analog-to-digital conversion is time- and energy-consuming, limiting the speed and efficiency of the neural network's performance. Photonic computing, which uses analog light signals, is one of the most promising approaches to addressing the issue.
In the aforementioned new study, the researchers designed an integrated photoelectronic processor to harness the advantages of both — light in the form of photons, and electrons as found in electric currents — in an all-analog way. This resulted in what is called an “all-analog chip combining electronic and light computing,” or ACCEL.
Tests showed that ACCEL is able to recognize and classify objects with a degree of accuracy comparable to that of digital neural networks. Furthermore, it classifies high-resolution images of various scenes from daily life over 3,000 times faster, consuming an astounding 4,000,000 times less energy than a top-of-the-line graphics processing unit (GPU).
Dai Qionghai, director of the School of Information Science and Technology at Tsinghua University, said that the team has developed a prototype chip, and will work toward making a general-purpose artificial intelligence chip for a broader range of applications.