Chinese researchers have developed the world's first neural dynamical system chip based on phase-change memristors, reducing single-step computation latency to 2.12 milliseconds and outperforming current advanced graphics processors by 50 to 478 times in brain cortex reconstruction tasks.

The breakthrough, published in Science on Friday, was led by Professor Yang Yuchao from Peking University's School of Integrated Circuits, who told Guangming Daily that enabling machines to model and understand the physical world in real time requires a "neural dynamical system" that combines neural networks with differential equations. Such systems can reconstruct smooth, precise 3D brain structures from incomplete and noisy data, with vast application potential.

However, traditional computing architectures suffer from a core bottleneck: the separation of memory and computation. During the solving process, massive intermediate variables shuttle repeatedly between memory and processors, like a huge data factory wasting time on transportation, resulting in huge latency and high power consumption.

To address this, the team exploited a unique physical property of phase-change memory – "conductance drift," which is predictable and precisely controllable within a certain time window. Based on this, they proposed a new paradigm called "controllable in-memory computing." In simple terms, what used to require repeated digital computations, cache access and data movement is now carried out by the device's own physical evolution.

In addition, the team mapped neural network weights onto multilevel conductance states of the phase-change memory, enabling matrix multiplication and accumulation within the same array. These two core computing tasks were thus integrated into a memory-computing array with a total area of just 0.28 square millimeters, fabricated using a 40-nanometer process. The chip achieves a single-iteration latency of 2.12 milliseconds, marking the first time neural dynamical hardware has entered the millisecond era.

Yang said the chip outperforms current state-of-the-art dedicated accelerators by 3.82 to 36.27 times in speed while significantly reducing power consumption. In brain cortex surface reconstruction tasks, it is up to 478.18 times faster than advanced foreign GPUs. The reconstructed cortical meshes are smooth, topologically consistent, accurately capture complex folding structures, and effectively suppress artifacts and self-intersections common in traditional methods.

Yang added that this breakthrough opens up new possibilities for brain-computer interfaces and brain disease diagnosis. In the future, individualized and dynamic brain digital twins may become feasible, providing a real-time hardware foundation for intraoperative neural navigation, early screening for Alzheimer's disease and personalized intervention.