https://j.people.com.cn/n3/2026/0708/c95952-20475568.html

https://www.science.org/doi/10.1126/science.aee6277

A research team led by Professor Yang Yuchao from the School of Integrated Circuits at Peking University and Researcher Song Zhitang from the CAS Shanghai Institute of Microsystem and Information Technology succeeded in developing a neurodynamics system chip using phase-change memristors. They shortened the per-step delay in complex calculations to 2.12 milliseconds, and for tasks like reproducing the brain cortex, the chip achieved speeds 50 to 478 times faster than the current advanced GPUs.

The research team utilized a unique property of phase-change memory called “conductance drift,” meaning the change in conductance over time can be predicted and precisely controlled. Based on this, they proposed a new computing paradigm called “controllable in-memory computing.” Simply put, work that previously required complex digital circuits to repeatedly perform tasks like calculations, cache access, and data transfer can now be carried out according to the physical laws of the devices themselves.

Another noteworthy point is that the team mapped the weights of a neural network to the multi-level conductance states of phase-change memory and simultaneously performed matrix multiplications within the same array. As a result, two major computational tasks were integrated into a memory-compute hybrid array with a total area of just 0.28 square millimeters. Manufactured with a 40-nanometer process, this chip achieved a delay of 2.12 milliseconds per iteration, bringing neurodynamics hardware into the millisecond era for the first time.