The Chinese Chip 478 Times Faster than NVIDIA A100 Virtually Clones the Brain in Real Time
The Chinese Chip 478 Times Faster than NVIDIA A100 Virtually Clones the Brain in Real Time
A group of researchers from Peking University and the Chinese Academy of Sciences has developed a new 40 nm chip designed for real-time reconstruction of the cerebral cortex. The device directly integrates an artificial intelligence into the hardware and completes the reconstruction of the brain's complex folded surface in less than half a second, a result that could expand possibilities in clinical settings and neuroscientific research.
According to the authors of the study published in the journal Science, the new processor achieves performance from 50 to 478 times greater compared to systems based on the NVIDIA A100 GPU for this specific workload. The advantage comes from adopting a computing-in-memory architecture, which combines memory and processing in the same array, eliminating much of the latency caused by the continuous data transfer between memory and processor.
The researchers also leveraged a characteristic of phase-change memristors usually viewed as a limitation, known as conductance drift. Instead of compensating for this phenomenon, the team employed it to perform calculations related to neural network dynamics, aiming for faster processing while also being more energy-efficient.
The reconstruction of the brain's surface represents one of the most challenging operations in neurological imaging. The numerous folds of the cortex increase the available surface area to accommodate billions of neurons but make the calculations required to obtain an accurate three-dimensional model extremely complex. Traditional systems require very powerful hardware and long processing times, factors that limit the use of these technologies when an immediate response is needed.
According to Yang Yuchao, the lead author of the research and a professor at the School of Integrated Circuits at Peking University, the new chip accurately reconstructs the cerebral cortex and could provide a hardware foundation for numerous medical applications. These include brain-computer interfaces, diagnosis and treatment of neurological diseases, real-time intraoperative neuronavigation, early Alzheimer’s screening, and personalized therapeutic interventions.
“In the future, it will be possible to create customized and dynamic digital brain twins,” Yang explained. The system also “provides a hardware foundation capable of operating in real time for intraoperative neuronavigation, early Alzheimer’s screening, and personalized interventions.”
An analysis conducted by researchers from the Juelich Research Centre also highlights the advantages of the computing-in-memory approach. Experts explain that processing data directly at the point of storage significantly reduces response times and energy consumption.
The platform offers high-fidelity calculations with latency on the order of milliseconds and could find applications not only in clinical imaging but also in robotics and embodied artificial intelligence systems. The authors also point to the possibility of using the chip for real-time monitoring of the cortical surface during neurosurgical procedures, directly supporting clinical decision-making processes.