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

August 20, 2026 (MLN): A Chinese-led research team has developed a vision chip capable of converting light directly into AI tokens, eliminating several conventional data-processing stages and significantly reducing energy consumption.

The technology, developed jointly by researchers at Nanjing University and the National University of Singapore, was detailed in a study published Wednesday in Nature Sensors, as reported by APP.

Unlike conventional image sensors that capture light before digitizing, buffering, transferring and processing the data, the new chip performs token generation directly within the physical device.

The chip incorporates an array of specialized pixels that can simultaneously detect light, store information locally and perform computations. Peripheral circuits can activate selected pixel regions, dividing an image into individual blocks for processing.

Once a block is selected, a sequence of voltages is applied to the pixels, allowing the stored optical information to undergo mathematical operations directly in the analog domain. The resulting current is then used as the output token.

Researchers described the process as “physical tokenization,” transforming the chip from a conventional image-capturing device into a system capable of generating visual-semantic information.

Testing showed that a system using the chip achieved 87.3% accuracy in image recognition, approaching the performance of conventional software-based methods. The tokenization process also delivered more than 10-fold higher energy efficiency.

The researchers said the approach removes data movement between sensing, memory and computing stages, addressing a major source of energy consumption in AI hardware.

The technology could pave the way for lower-power, near-zero-latency vision systems for applications including embodied intelligence, drones and smart security, while helping overcome computing and energy constraints facing edge AI.

Copyright Mettis Link News