Chinese Researchers Develop LightTok Chip to Cut AI Energy Consumption

Researchers in China have engineered a new sensor technology capable of significantly reducing the energy demands of drones and other autonomous systems during visual data processing. The study, published on August 19 in the journal Nature Sensors, introduces a two-dimensional chip named “LightTok” that transforms raw light into AI tokens directly on the sensor, bypassing several power-intensive conventional steps.

The design concept originated from Miao Feng, director of Nanjing University’s Institute of Brain-Inspired Intelligence. “Our design idea was to move token generation onto the sensor itself, allowing the chip to directly produce tokens that AI models can process once light reaches the sensor,” Miao stated. He emphasized that these tokens retain complete image information necessary for accurate interpretation.

Traditional visual perception systems require light signals to pass through multiple stages: capture by a sensor, conversion to digital pixels via analog-to-digital converters, temporary storage, and transfer to a separate processing unit where images are divided into grid-like compartments for tokenization. Research indicates that analog-to-digital conversion alone accounts for approximately 66% of an image sensor’s energy use. Furthermore, transmitting this data to the cloud increases overall power consumption.

LightTok addresses these inefficiencies by integrating sensing, memory, and computation into a single pixel. The chip utilizes an array of single-layer molybdenum disulfide floating-gate phototransistors. Molybdenum disulfide is a 2D material sensitive to light that can be produced in sheets just one atom thick. The phototransistor converts incoming photons into electrical current, while a floating gate within the component traps electrical charge, preserving data even after the light source is removed.

“The chip physically eliminates data movement, which is the main source of energy waste,” explained Liang Shi-Jun, a physics professor at Nanjing University, in remarks cited by Xinhua. This architecture allows light to enter and tokens to exit directly, giving the device its name.

In testing, the LightTok chip achieved an image recognition accuracy of 87.3%, comparable to traditional multi-step processes, while demonstrating tenfold greater energy efficiency in converting light into tokens.

Currently, the chip’s resolution is limited to 32 by 32 photosensitive pixels, a fraction of the quality found in modern smartphone cameras or drone sensors. However, the researchers believe the technology can be scaled up using complementary metal-oxide-semiconductor manufacturing processes, the standard method for producing chips in consumer electronics and autonomous hardware. Successful scaling could revolutionize remote sensing, enabling drones to extend flight times by reducing the energy required for visual processing.

Kumar Sokka, CEO of Acre Security, described the work as a “small-scale demonstration” but noted its significance for the physical AI sector. Sokka highlighted that the industry often focuses excessively on AI models rather than the energy cost of converting sensor data into usable formats at the point of detection. While he acknowledged that processing at the sensor level is a clever enabler for pervasive physical AI, he cautioned that it is not a complete solution to the broader energy challenges in the field.