Chinese Smart Chip Reconstructs Brain Structure Up to 478x Faster Than NVIDIA A100 GPU

Chinese Research Team Develops Breakthrough Brain Imaging Chip

Chinese researchers have developed a memory chip capable of simulating complex brain structures in real time, with the team stating the technology could improve diagnosis of neurological conditions, advance brain-computer interface applications, and support surgical navigation.

The 40nm chip was co-developed by researchers from Peking University and the Chinese Academy of Sciences, integrating artificial neural networks directly into hardware. According to the research team, the chip can reconstruct the brain's complex folded cortical surface in under half a second.

Researchers noted that when performing this task, the chip operates 50 to 478 times faster than systems equipped with NVIDIA A100 GPUs. The performance gains stem from a computing-in-memory (CIM) architecture — which integrates data storage and computation within the same memory array — dramatically reducing the latency caused by shuttling data between separate memory units and processors in conventional systems.

Rather than treating conductance drift in phase-change memristors as a defect, the team ingeniously leveraged the phenomenon to perform neural dynamic computations, enabling fast and energy-efficient processing.

Merging Memory and Computation

The chip was designed to address a long-standing bottleneck in brain imaging: the enormous data volumes required to reconstruct the brain's highly folded cortical surface are difficult for conventional hardware to handle. Faster processing could make advanced brain modeling more practical in hospital settings, where clinicians often need real-time results to support diagnostic and treatment decisions.

Yang Yuchao, a professor at the School of Integrated Circuits at Peking University and associate dean of the School of Electronics and Computer Engineering, is the lead author of the study. He stated that the chip is capable of accurately reconstructing the brain's folded cortex for medical applications.

Speaking to state-run outlet Guangming Daily, Yang said: "This breakthrough opens up entirely new possibilities for brain-computer interfaces and the diagnosis and treatment of brain diseases."

He added: "In the future, personalized, dynamic digital brain twins will become possible." Yang also noted that the technology "provides real-time-capable hardware for intraoperative neural navigation, early Alzheimer's disease screening, and personalized interventional therapy."

The human brain contains intricate folded structures whose grooves increase cortical surface area, allowing billions of neurons to be packed within the skull. Traditionally, reconstructing these structures requires powerful computing systems and lengthy processing times, limiting their applicability in time-sensitive clinical environments.

Accelerating Clinical Brain Imaging

The new design eliminates one of the biggest bottlenecks in conventional computer architectures — the physical separation between memory and processor. By integrating both functions on a single chip, the system reduces both power consumption and latency.

Researchers at the Jülich Research Centre in Germany, in an accompanying analysis piece, likened the approach to "processing raw milk directly at the dairy farm rather than transporting it to a factory" — highlighting the efficiency advantages of performing computation at the site of data storage.

They wrote that the platform enables "high-fidelity computation with millisecond-level latency," opening pathways for real-time applications in clinical imaging, robotics, and embodied intelligence. The researchers also stated that the work "holds promise for real-time tracking of cortical surfaces during neurosurgical procedures and could be integrated into clinical decision-making workflows."

The study has been published in the peer-reviewed journal Science.

原文來源： 查看原文