China's New AI Chip Maps the Human Brain Up to 478 Times Faster Than Nvidia's A100 GPU
Image Courtesy: Getty ImagesChinese researchers have developed a specialized memory chip that can reconstruct the brain’s complex folded surface in real time, a breakthrough that could accelerate medical. . .
Image Courtesy: Getty ImagesChinese researchers have developed a specialized memory chip that can reconstruct the brain’s complex folded surface in real time, a breakthrough that could accelerate medical imaging, brain-computer interfaces, and surgical navigation. The chip, built by researchers from Peking University and the Chinese Academy of Sciences using a 40-nanometer process, integrates an artificial neural network directly into its hardware. The team says it can reconstruct the brain’s intricate cortical folds in less than half a second, dramatically reducing the time needed for a computationally demanding task. The findingswere published inScience. The researchers reported that the chip performs between 50 and 478 times faster than systems powered by Nvidia’s A100 GPU for this specific workload. Its speed comes from a computing-in-memory architecture, which combines data storage and computation within the same memory array. This eliminates much of the delay caused by constantly moving data between separate memory and processing units. Instead of treating “conductance drift” in phase-change memristors as a hardware flaw, the team repurposed the characteristic to perform neural dynamical computations. This approach enables faster processing while reducing energy consumption compared to conventional computing architectures. Lead author Yang Yuchao, a professor at Peking University’s School of Integrated Circuits, said the technology could support real-time neuronavigation during surgery, early Alzheimer’s disease screening, and personalized treatments. He also suggested it could pave the way for dynamic digital brain twins tailored to individual patients. The human brain’s folded cortex dramatically increases its surface area, allowing billions of neurons to fit inside the skull. Reconstructing these structures has traditionally required powerful computing hardware and lengthy calculations, making real-time applications difficult in clinical settings. Researchers from Germany’s Juelich Research Centre, who provided an accompanying analysis, compared the architecture to processing raw milk directly on a dairy farm instead of transporting it to a factory. They said the system delivers high-fidelity calculations with millisecond-scale latency and could support future applications in clinical imaging, robotics, and embodied intelligence.