BEIJING, Aug. 21 (Xinhua) -- A Chinese-led research team has developed a new type of vision chip that directly converts light into tokens, bypassing the multiple processing steps that typically consume power and slow down performance.
The study, jointly conducted by Nanjing University and the National University of Singapore, was published Wednesday in the journal Nature Sensors.
Unlike conventional sensors, which capture light and then digitize, buffer, and transfer data before token generation, the new device performs the conversion in the physical domain.
It features an array of specialized pixels, each capable of simultaneously sensing light, storing data locally, and performing computations. The chip's peripheral circuits can selectively activate pixel regions to divide an image into blocks.
For each selected block, a specific voltage sequence is applied, enabling the stored optical information to carry out mathematical operations directly in the analog domain, with the output current serving as the token.
"The chip physically eliminates data movement, which is the main source of energy waste," said Liang Shijun, a professor at Nanjing University. "Light comes in, tokens come out."
This approach, described as "physical tokenization," turns the chip from a passive image recorder into an active visual-semantic generator, Liang said.
In tests, a system using the chip achieved 87.3 percent accuracy in image recognition, close to conventional software baselines, while improving energy efficiency by more than 10 times during the tokenization stage, he said.
"This design overturns the traditional 'sense-buffer-compute' sequence and opens a new path to break the power and computing bottlenecks of edge AI in the post-Moore era," said Miao Feng, a professor at Nanjing University.
He added that the technology could provide a highly efficient, near-zero-latency vision foundation for embodied intelligence, drones and smart security, accelerating the large-scale deployment of physical AI. ■
