Chinese researchers from Tsinghua University have unveiled the world's first all-analog photoelectronic chip, ACCEL, marking a significant leap forward in computer vision and AI technology. Published in the journal Nature, their research introduces a promising alternative to the energy-consuming analog-to-digital conversion process prevalent in existing technologies.
ACCEL, an abbreviation for "All-Analog Chip Combining Electronic and Light Computing," harnesses the synergistic advantages of light (in the form of photons) and electricity (through electric currents) in an all-analog manner. The chip's integrated photoelectronic processor enables it to process computer vision tasks at an unprecedented speed and energy efficiency.
Tests conducted on ACCEL have exhibited its exceptional capabilities, rivaling those of digital neural networks in object recognition and classification tasks. Remarkably, ACCEL outperforms top-of-the-line graphics processing units (GPUs) by processing high-resolution images of daily life scenes more than 3,000 times faster while consuming an astonishing 4,000,000 times less energy.
Analog and digital signals are two distinct types of signals that carry information. Analog signals, such as the rays of light forming an image, vary continuously, while digital signals, like binary numbers, are non-continuous.
In vision-based computing tasks like image recognition and object detection, signals from the environment are in analog form and must be converted into digital signals for processing by AI neural networks. However, the analog-to-digital conversion process is time-consuming and energy-intensive, limiting the overall speed and efficiency of neural network performance. The Tsinghua team's all-analog approach, leveraging analog light signals in photonic computing, addresses these challenges.
Fang Lu, a researcher from the Tsinghua team, explains that they have maximized the advantages of light and electricity under all-analog signals, effectively bypassing the drawbacks of analog-to-digital conversion and breaking the barriers of power consumption and speed.
Nature editors have lauded the Tsinghua research team for their innovative approach, minimizing the need for energetically costly analog-to-digital converters. Describing it as a "refreshing and pragmatic approach to artificial-intelligence hardware," they recognize ACCEL's high energy efficiency and commend its utilization of both electronic and photonic computing technologies.
The ultra-low power advantage of ACCEL holds great promise for addressing the heating issues associated with chip scaling. This breakthrough paves the way for future chip designs that are not only more energy-efficient but also capable of delivering unparalleled performance.
Dai Qionghai, director of the School of Information Science and Technology at Tsinghua University, reveals that the team has already developed a prototype chip. Their future endeavors involve creating a general-purpose artificial intelligence chip to cater to a broader range of applications, opening doors to transformative possibilities in fields such as healthcare and autonomous vehicles.