Peking University Develops the World's First Neurodynamics Chip on a 40nm Process
A Peking University team developed the world's first neurodynamics chip based on phase-change memristors on a 40nm process, compressing the computation step to 2.12 milliseconds with up to 478x speedup.
In an achievement that deepens China's presence in the AI chip race, a team from Peking University announced the development of the world's first "neurodynamics" chip based on phase-change memristors. The real achievement is not in size alone, but in speed: the chip succeeded in compressing the single-step computation time for complex neurodynamic systems to just 2.12 milliseconds, in what is described as breaking a computational bottleneck that had stood for nearly half a century. The research was published in the journal Science on July 3, 2026.
What Exactly Did the Team Develop?
The project was led by Professor Yang Yuchao of the School of Integrated Circuits at Peking University, in collaboration with the team of Professor Song Zhitang of the Shanghai Institute of Microsystem and Information Technology (under the Chinese Academy of Sciences). The chip uses a controllable "in-memory computing" architecture, is manufactured on a 40-nanometer process, and its computing array occupies no more than 0.28 square millimeters, with an operating frequency of 50 MHz. It thus combines memory and processing in one place, bypassing the "von Neumann bottleneck" that separates them in traditional computers and wastes energy shuttling data back and forth.
Why "40 Nanometers" Is Not the Story
The "40 nanometer" figure may seem modest compared to smartphone chips that have reached 3 nanometers, and therein lies the point. The achievement is not in miniaturization but in the architecture itself. China is constrained by export controls that deny it the latest ultra-precise chip-manufacturing equipment, so this work came to prove that a leap in performance is possible by redesigning how computation is done, rather than chasing the smallest transistor. In short: an architectural advantage that works around manufacturing constraints instead of colliding with them.
Neurodynamics: Harder Than Simulating Neurons
Most "neuromorphic" chips focus on simulating neuron spiking. "Neurodynamics" is harder: it is solving the differential equations that describe how the brain's state evolves over time, dense computations that had remained too slow for real-time applications. Reducing the single-step time to the millisecond level means, for the first time, the possibility of performing this simulation in real time, opening the door to applications that were practically impossible.
The Numbers and Potential Applications
In tasks like 3D reconstruction of the brain's cortical surface, the study says the chip achieved a speedup ranging from 50 to 478 times compared to the most powerful current GPUs, with an energy-efficiency improvement measured in tens to hundreds of times. This kind of low-power, real-time performance makes the chip a promising candidate for brain-computer interface applications, smart prosthetics, and wearable medical devices, where instant response and low consumption are decisive conditions.
A Balanced Reading: A Research Achievement, Not a Commercial Product
Despite the news's importance, it should be read carefully. What was achieved is a prototype that proved feasibility in specific tasks published in a scientific paper, not a chip ready for wide industrial production. The impressive numbers (50–478 times) pertain to particular neurodynamics tasks, not all types of computing. And the road from a university lab to a mass-manufactured commercial product is long and complex, as we have seen with previous neuromorphic chips that stumbled between research and market. The real value here is proving that innovative architecture can compensate for manufacturing constraints, a message as strategic as it is technical.
This work remains an indicator of a broader trend: as global technological competition intensifies and export controls tighten, Chinese research centers are turning to innovating in "how to compute" rather than in "transistor size." Whether this prototype turns into a product or remains a research achievement, it reshapes the contours of next-generation chip competition, where the architecture, not manufacturing precision alone, may become the real battleground.