Summary: Researchers have unveiled the world's first chip that can match the processing speed of the human brain.
Built using a standard 40 nanometer process, the sub-10-millisecond neural dynamical system uses phase-shift memristors to perform basic math calculations within memory. Occupying just 0.28 square millimeters, the chip achieves 478 times the speed over enterprise GPUs in real-time reconstruction of the cortex while significantly reducing power demands.
This milestone paves the way for real-time brain-computer interaction, intraoperative navigation, and full-scale digital brain twins.
Important Facts
- Sub-10-Millisecond Brain-Speed Simulation: The chip processes continuous neural dynamics at a processing speed that matches the native millisecond time scale of the human brain.
- Maximum GPU acceleration: In the works of 3D cortical surface reconstruction (white brain mapping and gray matter), a phase shift memristor chip has been achieved speed 478.18× compared to the enterprise-class NVIDIA A100 GPU.
- Superior Energy and Latency Metrics: Compared to modern Application-Specific Integrated Circuits (ASICs), the neuromorphic architecture works. 3.82× to 36.27× faster while eating 11.75× to 24.73× minimum power.
- Overcoming the Memory Wall: By using a memory computer with 9 stages of pipelines running at 50 MHz, the system eliminates the traditional data deadlock between memory and CPU/GPU processors.
- High-Fidelity Anatomical Reconstruction: The system generated smooth, closed, and topologically accurate 3D multicortical meshes, scoring high on the Average Symmetric Surface Distance (ASSD) and Hausdorff Distance metrics for neuroimaging accuracy.
Source: Peking University
A research team led by Professor Yang Yuchao from Peking University, and researchers from the Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, has created the first chip in the world that can keep up with the speed of the human brain.
The study, titled “A sub-10-millisecond neural dynamical system based on phase shift memristors,” was published in Science.
It's the background
Neural dynamical systems combine neural networks with mathematical models that describe how complex systems change over time. They are useful for physical modeling, medical imaging, and three-dimensional brain reconstruction. However, these systems require repeated calculations, error checking, and size adjustments for each calculation step. In conventional computers, data must also move frequently between memory and the processor, increasing processing time and power consumption.
Why is it important
Fast and accurate brain modeling is essential for technologies that must respond in real time, including brain-computer interfaces, surgical navigation, and medical imaging. Existing hardware often requires a lot of time and energy for these necessary calculations. By performing critical tasks directly in memory, the new chip minimizes data movement and brings high-quality brain modeling closer to real-time use.
Key findings
Built using a 40-nm process, the in-memory computer chip and conductance-drift arrays take up only 0.28 square millimeters. It operates at 50 MHz and uses nine pipeline stages for each integration step. For neural dynamics calculations, the system is 3.82 × to 36.27 × faster and consumes 11.75 × to 24.73 × less power than state-of-the-art ASICs, or application-specific integrated circuits. In cortical surface reconstruction tasks, it reaches 478.18× speedup compared to NVIDIA A100 GPU.
The researchers used the chip to reconstruct the brain's white and gray matter and generate 3D-based meshes in real time. The system produced smooth, closed, and topologically consistent column surfaces while accurately capturing complex brain folds. It also performed well on the average distance and Hausdorff distance estimates, demonstrating its ability to support reliable brain modeling.
Implications for the Future
The chip can help move complex neural modeling from slow, offline processing to millisecond-scale operations. In the future, the technology may support brain-computer communication, digital brain twins, real-time surgical navigation, brain surface reconstruction, and tools to study neurodegenerative diseases such as Alzheimer's and Parkinson's.
Important Questions Answered:
A: A phase shift memristor is a memory element that changes its electrical resistance when exposed to body temperature conditions, allowing it to store continuous analog values rather than binary 0's and 1's. This allows the chip to calculate complex mathematical calculations directly within the memory cluster, skipping time-consuming data transfers.
A: Conventional GPUs process heavy physical and mathematical models by constantly sending data back and forth between computing units and high-bandwidth memory. A phase-shifting chip performs these calculations locally, achieving 478 times the speed of 3D collar reconstruction over the A100 GPU while drawing half the power.
A: Millisecond-scale processing enables real-time computer interfaces (BCIs) that interpret neural signals without lag, surgical navigation systems that stimulate 3D brain scans during procedures, and real-time “digital brain twins” to mimic neurodegenerative disorders such as Parkinson's and Alzheimer's.
Editor's Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper is fully revised.
- More content has been added by our staff.
About this neurotech research news
Author: Jiang Zhang
Source: Peking University
Contact person: Jiang Zhang – Peking University
Image: Image posted in Neuroscience News
Actual research: Open access.
“A sub–10-millisecond neural dynamical system based on phase shift memristors” by Lei Cai, Yaoyu Tao, Chenchen Xie, Longhao Yan, Shiqian Li, Ruihong Shen, Zelun Pan, Xile Wang, Bowen Wang, Daijing Shi, Yihang Zhu, Teng Zhang Xie, Ying Lixin, Zhuang Lixin, Yuxin Yang. Science
DOI:10.1126/science.aee6277
Abstract
A sub-10-millisecond neural dynamical system based on phase-shift memristors
High-reliability geometry for physical earth modeling requires real-time, dense, and separable wear fields on the manifold.
Neural dynamical systems (NDSs) using a combination of dynamic step size and embedded neural networks excel in these tasks but still suffer from delays on the order of hundreds of milliseconds. In this work, we report a sub-10-millisecond NDS hardware that uses precisely controlled driving of phase-shift memristors and their computing capability for multi-level memory.
We developed a 40 nanometer NDS chip to perform the task of reconstructing a challenging environment. Compared to state-of-the-art NDS hardware, our NDS design achieves a latency of 2.12 milliseconds (less than 10 milliseconds) for multiple NDS calculations and a fault tolerance of 10.-7 and delivers 3.82× to 36.27× speed while consuming 11.75× to 24.73× less power.
The end-to-end latency of NDS using hardware benchmarks and simulations exceeded that of the A100 graphics processing unit by 50.38× to 478.18×.
