Chinese team develops vision chip that turns light directly into AI tokens

A research team led by scientists in China has built a vision chip that converts light straight into AI tokens, skipping several standard data handling steps. The work was carried out by teams at Nanjing University and the National University of Singapore. The study was published on Wednesday in Nature Sensors, Mettis Global reported, citing APP.

Conventional image sensors record light, convert it into digital data, store it and then send it on for processing. The new design instead generates tokens inside the chip itself, using pixels that can sense light, hold information and run calculations. Selected regions of an image are handled as separate blocks, with voltages applied so that the stored optical data undergoes analog mathematical operations, and the resulting current serves as the output token.

In testing, a system built around the chip reached 87.3% accuracy in image recognition, close to software-based methods, and tokenization delivered more than 10-fold higher energy efficiency. The researchers said the design removes data movement between sensing, memory and computing, a major source of energy use in AI hardware. They suggested it could support low-power vision systems for drones, smart security and embodied intelligence.