Chinese Researchers Develop New Chip for Millisecond-Level Neural Dynamics
Chinese researchers have developed the world's first neural dynamics system chip based on phase-change memristors. The technology allows computations to be performed within milliseconds.
Researchers from Peking University and the Shanghai Institute of Microsystem and Information Technology at the Chinese Academy of Sciences have announced a significant breakthrough in neural dynamics computing. Their work, published in the prestigious journal Science, details the creation of the world's first neural dynamics system chip utilizing phase-change memristors, capable of operating at millisecond timescales.
The research addresses a long-standing challenge in employing phase-change memristors: enabling and controlling "in-memory computing" operations. The new chip reportedly achieves the unprecedented feat of compressing the single-step computation latency of a neural dynamics system to just 2.12 milliseconds.
Neural dynamics systems combine the expressive power of neural networks with the continuous evolution mechanisms of differential equations, holding potential for applications in physical world modeling and computational imaging. Historically, achieving low-latency, real-time computation while maintaining high-precision continuous modeling has been a major bottleneck for practical deployment.
The team introduced a novel paradigm for "controllable in-memory computing" by precisely regulating the phase-change memristors' properties. This allows for efficient computation directly within the memory, synchronizing device physics with neural dynamics algorithms. The chip, manufactured using a 40nm process, occupies a mere 0.28 square millimeters and integrates all necessary circuitry, operating at 50 MHz with a nine-stage pipeline for each integration step.
Experimental results demonstrate that the new chip significantly outperforms current state-of-the-art specialized accelerators (ASICs) and even NVIDIA A100 GPUs in speed and energy efficiency. For applications like brain-computer interfaces (BCIs), the rapid millisecond-level brain modeling could be transformative, potentially enabling real-time analysis of neural signals, prediction of brain states, and closed-loop control, advancing BCI from simple signal recognition to intelligent brain modeling and interaction.