Chinese scientists at the University of Science and Technology of China in Hefei, Anhui province, have made a significant breakthrough in AI technology. The computer chip known as ACCEL (an all-analog chip combining electronic and light computing), has the potential to outperform Nvidia’s A100 by a remarkable 3,000 times in AI tasks.
The Chinese researchers have been working on developing PIC-based AI chips for several years. In 2021, they developed a PIC-based AI chip that was 100 times faster than the A100. The new chip is a significant improvement over the previous chip, and it represents a major breakthrough in the field of PIC-based AI computing.
The researchers published a paper describing their work in the journal Nature Photonics on November 1, 2023.
This game-changing development is rooted in the innovative concept of photonic computing, which leverages light for both data extraction and calculations, as opposed to traditional semiconductor chips that rely on electricity.
ACCEL’s power lies in its ability to harness the speed and energy efficiency of light. In laboratory tests, it demonstrated a staggering capability to perform image recognition tasks at a speed 3,000 times faster than the A100 while consuming a mere fraction of the energy – an astonishing 4 million times less. To put this into perspective, ACCEL can execute an impressive 74.8 quadrillion operations per second using just a single watt of power, making it a frontrunner in energy efficiency.
Beyond speed and efficiency, ACCEL streamlines the processing of data, reducing the time required to handle each frame to a mere 72 nanoseconds. It eliminates the need for power-intensive analog-to-digital converters, enhancing its practicality.
The researchers say that the chip could be used for a variety of AI applications, including image recognition, natural language processing, and autonomous driving. They also say that the chip is scalable, meaning that it can be made even faster and more powerful in the future.
Unlike the general-purpose chips found in common devices like smartphones and computers, this chip doesn’t possess versatility across a wide range of functions.
The scientists at Tsinghua University acknowledge the significance of this achievement, calling it a pinnacle in developing a new computing architecture for the AI era. They highlight the pressing challenge of transitioning this breakthrough into practical applications that address significant national and public needs.
The applications of the ACCEL chip are promising and wide-ranging. It holds potential in various AI domains, such as image recognition, natural language processing, and autonomous driving. Additionally, its scalability suggests the possibility of even more formidable capabilities in the future, potentially reshaping the landscape of AI computing.
