A Nanjing University team led by professors Miao Feng and Liang Shijun, working with collaborators from the National University of Singapore, has developed an intelligent visual sensor called the Guangyu Chip, according to a report in Science and Technology Daily. The outlet states that the device directly converts optical signals into tokens for a visual-model encoder, and that the work was published online on August 19 in Nature Sensors.
Science and Technology Daily describes the chip’s core as a photosensitive memory array built on monolayer molybdenum-disulfide floating-gate optoelectronic transistors. In the results conveyed by the outlet, the team reported 87.3% recognition accuracy and more than tenfold higher energy efficiency at the tokenization stage compared with a conventional approach. The report presents these figures as research outcomes tied to the described system and the specific comparison scope.
How the Guangyu Chip Converts Light Into Tokens
As described by Science and Technology Daily, the Guangyu Chip is designed to output tokens that a visual-model encoder can consume, rather than emitting only pixel data. In that framing, token generation occurs where light is captured, aligning sensing with the format a model expects. If such a front end reduces the need for intermediate conversions, it could simplify the handoff into downstream models, though any broader gains would depend on the rest of the system.
The reported implementation centers on a photosensitive memory array based on monolayer molybdenum-disulfide floating-gate optoelectronic transistors. The outlet characterizes this materials platform as the enabling substrate for directly mapping optical input to token output. By combining photosensitivity with a memory array, the device, as presented, integrates the capture of light with an output geared for a model encoder. Interpreted at a high level, this suggests a design that aims to bring early-stage representation closer to the tokens a model consumes, which could reduce data movement and preprocessing overhead in some scenarios. That interpretation is conditional and would be shaped by task requirements and system integration.
The publication venue noted by the outlet places the disclosure in Nature Sensors. That timing and venue indicate the work is positioned in a research context. Any implications for deployment would require separate validation beyond what is described in the report.
The research also adds a Nanjing angle to EastFrontier’s earlier report on AI chip scientist Song Yuhang joining Nanjing University.
Reported Accuracy and Tokenization Energy Metrics
According to Science and Technology Daily, the research team reported 87.3% recognition accuracy. The article treats this as a headline indicator of how the chip’s token output integrates with a visual-model encoder under the study’s conditions. While the figure offers a concrete reference point, it is tied to the evaluations described by the researchers, so readers should consider it within that defined experimental frame.
On energy use, the outlet cites a comparison scoped to the tokenization stage of the vision pipeline. It reports that the team measured more than a tenfold gain in energy efficiency at that stage relative to a conventional approach. The summary does not extend that comparison to the rest of the system, and it confines the claim to tokenization. As a result, any interpretation of end-to-end benefits would be speculative without further data. Still, if tokenization is a significant contributor to overall cost in a given application, a stage-specific improvement of this magnitude, as reported, could be meaningful for system design, provided the rest of the stack supports it.
The outlet attributes all performance figures to the research publication and notes their disclosure with the online appearance of the work in Nature Sensors on August 19. The reported numbers are therefore part of the study’s documentation and should be interpreted as such.
Collaboration, Materials Choice, and Publication Details
Science and Technology Daily credits the project’s leadership to professors Miao Feng and Liang Shijun of Nanjing University, and it notes the involvement of collaborators from the National University of Singapore. In presenting a cross-institutional effort, the report identifies the Guangyu Chip as the collective output of that team.
The materials and device structure receive specific emphasis in the article. It points to a photosensitive memory array realized with monolayer molybdenum-disulfide floating-gate optoelectronic transistors as central to the design. In the report’s depiction, that combination underlies the conversion of optical signals into tokens for a visual-model encoder. Interpreted from a systems perspective, anchoring token output in a photosensitive memory array could be a strategy for aligning sensing with representation, which may offer architectural pathways to limit data transformations early in the pipeline. That editorial context remains conditional, pending details beyond what the outlet summarizes.
The publication detail highlighted by the outlet is that the work appeared online on August 19 in Nature Sensors. This anchors the announcement to a specific date and venue, delineating the research status of the disclosure. The article refrains from characterizing the chip as commercially available, and it confines the reported outcomes to the recognition accuracy and tokenization-stage energy efficiency figures drawn from the study.
Taken together, the facts presented by Science and Technology Daily outline a research-stage vision chip that pairs a defined materials platform with a token-focused output for a model encoder. The team composition, the described device architecture, the reported accuracy figure, and the energy comparison limited to tokenization are the elements the outlet attributes to the research. Any broader performance claims, deployment timelines, or independent benchmarks would fall outside the scope of the report and are not asserted here.