Scientists in China have developed a powerful new chip that can perform complex brain modeling in less than 10 milliseconds, a speed comparable to the operating pace of the human brain.
The research was led by Yuchao Yang of Peking University, in collaboration with scientists from the Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences. The findings were published in Science.
The chip is designed for neural dynamical systems, which combine neural networks with mathematical equations to model how complex systems change over time. These systems are important in areas such as medical imaging, physical modeling, and three-dimensional brain reconstruction.
One major challenge is that reconstructing the brainโs folded surface requires enormous computing power. Conventional computers must constantly move data between memory and the processor, which slows computation and increases energy use.
The new chip solves this problem by performing key calculations directly where data is stored. This method, known as in-memory computing, reduces the need for repeated data transfer and allows the chip to work much faster while using less power.
Built using a 40-nanometer manufacturing process, the chip contains specialized memristor arrays that occupy only 0.28 square millimeters. It runs at 50 MHz and completes each integration step through nine pipeline stages.
During neural dynamics calculations, the chip performed 3.82 to 36.27 times faster than advanced application-specific integrated circuits. It also used 11.75 to 24.73 times less power.
In cortical surface reconstruction, the chip achieved a speedup of up to 478.18 times compared with an NVIDIA A100 GPU.
The researchers tested the chip by reconstructing the boundaries between the brainโs white matter and gray matter. It produced real-time three-dimensional surface meshes that were smooth, closed, and structurally consistent, while still preserving the brainโs complex folds.
The system also performed strongly on key accuracy measurements, including average symmetric surface distance and Hausdorff distance, which are used to compare reconstructed surfaces with target brain structures.
By moving complex neural modeling from slow offline computation to millisecond-scale operation, the technology could support future applications in brainโcomputer interfaces, surgical navigation, medical imaging, digital brain twins, and neurodegenerative disease research.
The breakthrough could also become useful in studying conditions such as Alzheimerโs disease and Parkinsonโs disease, where detailed and rapid brain modeling may help researchers better understand structural changes.
Overall, the study shows how memristor-based chips could bring high-speed, low-power computing closer to the way biological brains process information.
Journal Reference:
Cai, L., Tao, Y., Xie, C., Yan, L., Li, S., Shen, R., Pan, Z., Wang, X., Wang, B., Shi, D., Zhu, Y., Zhang, T., Zhu, Y., Li, X., Song, Z., Huang, R., & Yang, Y. (2026). A subโ10-millisecond neural dynamical system based on phase-change memristors. Science. https://doi.org/10.1126/science.aee6277