Chinese brain-mimicking chip outpaces NVIDIA GPU by up to 478x
Chinese researchers have developed a memory chip that can model complex brain structures in real time, a breakthrough they say could improve brain disease diagnosis, brain-computer interfaces, and surgical navigation. The 40-nanometer chip, created by researchers from Peking University and the Chinese Academy of Sciences, integrates an artificial neural network into its hardware. According to the team, it reconstructs the brain’s intricate folded surface in less than half a second. The researchers reported that the chip is 50 to 478 times faster than systems powered by Nvidia’s A100 graphics processing unit for this task. The performance gain comes from a computing-in-memory architecture that carries out data storage and computation in the same memory array, reducing delays caused by moving data between separate memory and processing units. Instead of treating “conductance drift” in phase-change memristors as a defect, the team used the property to perform neural dynamical computations, enabling fast and energy-efficient processing. The researchers designed the chip to address a long-standing challenge in brain imaging, where conventional hardware struggles to keep pace with the large volume of data needed to reconstruct the brain’s highly folded surface. Faster processing could make advanced brain modelling more practical in hospitals, where clinicians often need results quickly to support diagnosis and treatment decisions. Memory meets computation Lead author Yang Yuchao, a professor at Peking University’s School of Integrated Circuits and deputy dean of its School of Electronic and Computer Engineering, said the chip can accurately reconstruct the brain’s folded cortex for medical applications. “This breakthrough opens up new possibilities for brain-computer interfaces and the diagnosis and treatment of brain diseases,” Yang told state-run Guangming Daily. He added: “In the future, personalised and dynamic digital brain twins will become possible.” Yang also said the technology “provides a hardware foundation that can operate in real time for intraoperative neuronavigation, early screening for Alzheimer’s disease and personalised interventions.” The human brain contains complex folds that increase its surface area, allowing billions of neurons to fit inside the skull. Reconstructing those structures has traditionally required powerful computing systems and lengthy calculations, limiting their use in time-sensitive medical settings. Faster clinical brain imaging The new design removes one of the biggest bottlenecks in conventional computer architecture, where memory and processors are physically separated. By combining both functions on the same chip, the system reduces latency while lowering power consumption. In an accompanying analysis, researchers from Germany’s Juelich Research Centre compared the approach to “processing raw milk on a dairy farm instead of transferring it to a factory,” highlighting the efficiency of performing computation where data is stored. They wrote that the platform delivers “high-fidelity calculation with a millisecond-scale latency,” creating a path toward real-time applications in clinical imaging, robotics , and embodied intelligence. The researchers also said the work “could enable real-time cortical surface tracking during neurosurgery and could be integrated into clinical decision-making.” The study was published in the journal Science.
Source: Interesting Engineering