[News] Flexible AI Chip Integrating 10,628 Transistors Debuts, Paving the Way for Bendable Intelligence

Chinese researchers announced a major breakthrough on January 29 in a paper published in Nature, unveiling a fully flexible artificial intelligence chip that provides critical hardware support for applications such as wearable health monitoring devices and flexible robotics.

As artificial intelligence continues to converge with the Internet of Things and embodied intelligence, demand is rapidly rising for lightweight, efficient, and flexible intelligent computing hardware. Conventional silicon-based rigid chips struggle to conform to the human body or complex curved surfaces, while existing flexible processors are typically constrained by low operating frequencies, high power consumption, and limited parallel computing capabilities—making them ill-suited for data-intensive tasks such as neural network inference.

Researchers from Tsinghua University, Peking University, and other institutions have successfully developed the FLEXI series of fully flexible, digital compute-in-memory AI chips using domestic manufacturing processes, overcoming long-standing barriers that have limited the application of flexible electronics in high-performance edge AI computing. Based on low-temperature polycrystalline silicon thin-film transistors, the chip is ultra-thin, freely bendable, and features ultra-low power consumption, high energy efficiency, strong robustness, and low cost.

According to the research team, the chip adopts a compute-in-memory architecture centered on fully digital static random-access memory (SRAM), effectively integrating “memory units” and “computing units” into a single structure. This design eliminates the time and energy overhead associated with data movement, significantly improving computational efficiency. The smallest version, the FLEXI-1 chip, measures just 31.12 square millimeters and integrates 10,628 transistors, operating at an ultra-low power consumption of only 55.94 microwatts.

Yan Bonan, assistant professor at the Peking University Institute for Artificial Intelligence and one of the study’s authors, told reporters that the chip can withstand more than 40,000 cycles of 180-degree bending without performance degradation and has demonstrated stable operation over long-term tests lasting up to six months. It also supports neural network compression and one-click deployment, further enhancing on-chip intelligence. With a capacity of just 1 kilobit, a single flexible chip can achieve arrhythmia detection accuracy of up to 99.2%, positioning it as a core computing engine for next-generation wearable medical devices, flexible brain–computer interfaces, and intelligent robotic systems.

(Photo credit: FREEPIK)