Key Takeaways
- Record performance: The new chip achieves a latency of 2.12 milliseconds per iteration in dynamic reconstruction of the brain's surface, making it between 50 and 478 times faster than systems based on Nvidia A100 GPUs.
- Technology: The device uses phase-change memristors and exploits "conductance drift" for controllable in-memory computing, eliminating data transfer between memory and processor.
- Application impact: The chip cuts energy consumption by up to 24 times and opens the door to uses in surgical neuronavigation, early diagnosis of neurodegenerative diseases, and brain-computer interfaces.
A chip rebuilds the brain in real time
A team led by Professor Yang Yuchao of Peking University, in collaboration with the Institute of Microelectronics of the Chinese Academy of Sciences, has developed a chip capable of reconstructing the surface of the cerebral cortex in real time. The findings, published in Science, describe the device as the world's first hardware for dynamic neural systems based on phase-change memristors.
The bottleneck of von Neumann architecture
Traditional computers physically separate memory and processor. Every complex calculation requires constant movement of data between the two units, a process that consumes time and energy. This separation, known as the von Neumann bottleneck, has so far prevented real-time modeling of the complexity of the human brain.
Computing inside memory
The team harnessed conductance drift, a physical property of phase-change memristors previously regarded as a flaw, turning it into a computing tool. The resulting chip stores data while simultaneously performing mathematical operations within itself, following an approach defined as controllable in-memory computing. According to Yang, operations that once required digital cycles, cache access, and data transfers are now carried out by the physical evolution of the device itself.
The numbers behind the experiment
Built using a 40-nanometer manufacturing process, the chip reached a latency of 2.12 milliseconds per single iteration, marking the first entry of dynamic neural system hardware into the millisecond era. In tests reconstructing the cerebral cortex, it outperformed Nvidia A100 GPU-based systems by 50 to 478 times and, in some cases, beat specialized processors (ASICs) by up to 36 times. Energy consumption came in at up to 24 times lower than traditional solutions.
Clinical applications on the horizon
The ability to generate real-time three-dimensional brain maps opens the door to direct use in intraoperative neuronavigation, with potential benefits for precision during surgical procedures. The same principle could enable early screening for neurodegenerative diseases such as Alzheimer's, through continuous analysis of brain structures. The chip is also a candidate for accelerating the development of brain-computer interfaces and the creation of personalized digital twins of the brain—virtual models that could be used to test therapies and track the progression of a disease in individual patients.
