China has developed an ultra-compact computer chip that runs up to 478 times faster than Nvidia's graphics processing unit (GPU) systems. The chip is used to model complex brain structures in real time within 0.5 seconds. It is expected to revolutionize the diagnosis and treatment of brain disorders such as Alzheimer's and to enhance the performance of brain-machine interfaces. However, unlike Nvidia's GPUs, which are used across many fields, the chip is limited to brain science research.
Chinese scientists developed the chip, according to the South China Morning Post (SCMP) on Tuesday. The developers said the chip could not only revolutionize the diagnosis and treatment of conditions such as Alzheimer's disease but also enhance brain-machine interface performance and support surgical procedures.
Researchers from Peking University and the Chinese Academy of Sciences published the achievement as a peer-reviewed paper in the international journal Science. At its core is a 40-nanometer (nm) memory chip integrated with an artificial neural network. The device overcomes long-standing computational limits, reconstructing the complex brain surface within 0.5 seconds. The research team explained that it is 50 to 478 times faster than Nvidia's A100 GPU.
Yang Yuchao, a professor at Peking University's School of Integrated Circuits and the paper's lead author, told the state-run Guangming Daily that "this chip can accurately reproduce brain folds for medical use," adding that "this achievement opens new possibilities for brain-computer interfaces and the diagnosis and treatment of brain disorders." He continued, "In the future, personalized dynamic digital brain twins will also become possible," adding that "it also provides a hardware foundation to support neural navigation during surgery, early Alzheimer's screening, and personalized treatment in real time." A digital brain twin refers to a virtual model of a real person's brain replicated inside a computer.
The human brain is folded to increase its surface area. This is to fit billions of neurons within the physical limits of the skull. Previously, reconstructing these complex folds in real time required lengthy computation on expensive, large-scale equipment. In standard computer architecture, storage and processing units are separated, creating a bottleneck. As data moves back and forth between memory and processor, latency and power consumption increase.
The research team solved this problem by integrating a neural dynamics system into an "in-memory computing" architecture. In-memory computing performs data storage and computation within the same memory array, increasing speed and reducing power use.
They turned a flaw of phase-change memristors, a next-generation memory chip, to their advantage. Phase-change memristors are next-generation memory devices that can perform storage and computation simultaneously. Their drawback is a phenomenon called "conductance drift" that makes data unstable, but the team instead used it to quickly approximate state changes in complex brain models. Conductance drift is a phenomenon in which the position or structure of defects inside a memory cell changes, causing the stored conductance value to shift over time.
Two researchers from Germany's Jülich Research Center, who did not participate in the study, likened the approach of performing storage and computation in the same memory array to "processing crude oil directly at the oil field rather than transporting it to a factory," in a commentary in Science. They praised the platform, saying it "performs high-precision computation with millisecond (one-thousandth of a second) latency, opening the way to real-time computation in clinical imaging, robotics, and embodied intelligence."
