Chinese researchers have developed a new brain-inspired chip capable of reconstructing complex brain structures in less than half a second; in a specific brain-mapping task, the system outperformed an Nvidia 800 GPU by up to 478 times. The chip was developed by researchers from Peking University and the Chinese Academy of Sciences, and the results were published in the scientific journal *Science*.

The technology utilizes a 40-nanometer manufacturing process and was created to address a major challenge in modern neuroimaging. The human brain has an extremely complex surface. The cortex consists of a vast number of folds, sulci, and structures that must be digitally reconstructed from scan data—a process requiring immense computational effort. Traditional computers face an additional hurdle: memory stores information while the processor performs calculations, meaning massive amounts of data must constantly travel between these two system components.

That movement consumes energy and causes delays; Chinese researchers decided to mitigate this issue by using an architecture known as "processing-in-memory," a system in which storage and processing occur within the same structure. The chip employs devices called phase-change memory cells—components capable of storing information via different physical states—but the researchers took a particularly interesting approach.

A characteristic of these components known as "conductance drift"—normally viewed as a problem—was instead utilized as part of the computing process itself, enabling the chip to perform dynamic neural calculations directly within the memory. The result was a significant reduction in data processing time; in tests presented by the researchers, the system successfully reconstructed the brain's cortical surface in less than 500 milliseconds.

Depending on the configuration used, performance ranged from 50 to 478 times that of an Nvidia 800-based system for that specific task. This does not mean the new Chinese chip is hundreds of times more powerful than an H100 for any given application. The advantage arises precisely because its architecture was developed to execute that specific type of processing—mimicking the human brain.