Summary: Researchers have introduced the world’s first chip engineered to operate at the millisecond timescale of the human brain.
Built on a standard 40-nanometer manufacturing process, this sub-10-millisecond neural dynamical system uses phase-change memristors to perform core mathematical operations directly inside memory. The complete design fits within 0.28 square millimeters and delivers dramatic performance and energy-efficiency improvements—reaching up to a 478× speedup over enterprise GPUs in real-time cortical surface reconstruction while cutting power consumption by an order of magnitude. This hardware breakthrough brings real-time brain-computer interfaces, intraoperative surgical navigation, and full-scale digital brain twins much closer to practical use.
Key Facts
- Sub-10-Millisecond Brain-Speed Simulation: The chip runs continuous neural dynamics at the millisecond temporal resolution characteristic of biological brains.
- Massive GPU Acceleration: For 3D cortical surface reconstruction—the task of mapping white and gray matter folds—the phase-change memristor chip achieved up to a 478.18× speedup versus an enterprise NVIDIA A100 GPU.
- Superior Energy and Latency: Against leading application-specific integrated circuits (ASICs), the neuromorphic design is 3.82× to 36.27× faster while consuming 11.75× to 24.73× less energy.
- In-Memory Computing Eliminates the Memory Wall: With nine pipeline stages running at 50 MHz, the system removes the expensive data transfers between memory and processors that limit conventional architectures.
- High-Fidelity Anatomical Reconstruction: The chip produces smooth, closed, topologically accurate 3D cortical meshes and scores highly on metrics such as Average Symmetric Surface Distance (ASSD) and Hausdorff Distance.
Source: Peking University
A research team led by Professor Yang Yuchao from Peking University, in collaboration with colleagues at the Shanghai Institute of Microsystem and Information Technology (Chinese Academy of Sciences), developed this neuromorphic chip capable of matching human brain operating speed.
The study, titled “A sub–10-millisecond neural dynamical system based on phase-change memristors,” appears in the journal Science.
Background
Neural dynamical systems combine neural networks with differential equations to model how complex systems evolve over time. They are a powerful approach for physical simulation, medical imaging, and three-dimensional brain reconstruction, but they require repeated computations, adaptive step-size control, and frequent error correction. On conventional hardware, these computations are slowed by repeated transfers between processor and memory—a bottleneck often called the “memory wall”—which increases both latency and energy use.
Why this matters
Low-latency, high-accuracy brain modeling is essential for technologies that operate in real time, including brain-computer interfaces (BCIs), surgical navigation systems, and interactive medical imaging. Existing hardware often cannot meet the combined demands of speed, energy efficiency, and numerical precision. By performing crucial operations inside memory arrays using phase-change memristors, this chip substantially reduces data movement and enables neural dynamical systems to run at brain-like millisecond speeds.
Key findings
The chip was fabricated in a 40-nanometer process and integrates in-memory computing arrays that exploit precisely controlled conductance drift of phase-change memristors. The active area for these arrays is only 0.28 square millimeters. The design uses nine pipeline stages, running at 50 MHz, to execute each integration step of the neural dynamical system. Measured against state-of-the-art NDS hardware, this design delivers single-iteration latency as low as 2.12 milliseconds with tight numerical tolerance, yielding 3.82× to 36.27× higher speed while reducing power consumption by 11.75× to 24.73×. When compared end-to-end with an NVIDIA A100 GPU, the system demonstrates a 50.38× to 478.18× improvement in latency for cortical reconstruction workloads.
In practical tests, researchers used the chip to reconstruct cortical surfaces—generating smooth, closed, and topologically consistent 3D manifold meshes that faithfully capture the complex folds of white and gray matter. The results showed strong performance on quantitative metrics (ASSD and Hausdorff Distance), underlining the platform’s suitability for high-fidelity neuroimaging.
Future implications
This sub-10-millisecond neuromorphic system could shift complex neural modeling from offline, time-consuming computation to interactive, millisecond-scale operation. Potential applications include responsive brain-computer interfaces that process neural signals with minimal lag, intraoperative navigation systems that update three-dimensional brain images during surgery, and real-time digital brain twins for research into neurodegenerative conditions such as Alzheimer’s and Parkinson’s disease.
Key Questions Answered:
A: A phase-change memristor is a memory device whose electrical resistance changes according to the material’s phase state. It can store analog conductance values and perform computations directly where the data are stored. This compute-in-memory capability reduces costly data transfers and makes it feasible to run continuous, high-precision neural dynamical calculations at very low latency and energy cost.
A: Conventional GPUs must repeatedly move large datasets between processors and memory to evaluate complex physical and mathematical models. By contrast, the phase-change memristor chip computes in place inside memory arrays. For cortical surface reconstruction, that results in up to a 478× speed advantage over an A100 GPU while using a small fraction of the energy.
A: Millisecond-scale processing enables real-time BCIs that interpret neural signals with virtually no lag, intraoperative navigation systems that can refresh 3D brain models during surgery, and interactive digital brain twins that let clinicians and researchers simulate and study progressive neurological diseases in near real time.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full.
- Additional technical context was added by the editorial staff.
About this neurotech research news
Author: Jiang Zhang
Source: Peking University
Contact: Jiang Zhang – Peking University
Image: The image is credited to Neuroscience News
Original Research: Open access. “A sub–10-millisecond neural dynamical system based on phase-change memristors” by Lei Cai et al., published in Science. DOI: 10.1126/science.aee6277
Abstract
A sub–10-millisecond neural dynamical system based on phase-change memristors
Accurate geometry for real-world physical modeling requires dense, differentiable deformation fields on manifolds and low-latency computation. Neural dynamical systems (NDSs) that combine adaptive-step integration with embedded neural networks are well suited to these tasks but normally run with latencies of hundreds of milliseconds. This work demonstrates a sub–10-millisecond NDS hardware platform using phase-change memristors whose controlled conductance drift and multilevel compute-in-memory properties enable fast, precise integration.
A 40-nanometer NDS chip was fabricated and applied to challenging cortical surface reconstruction tasks. The measured single-iteration latency reached 2.12 milliseconds with an error tolerance of 10−7. Compared with state-of-the-art NDS hardware, the chip delivered 3.82× to 36.27× faster speed while using 11.75× to 24.73× less power. End-to-end comparisons to an A100 GPU showed improvements in latency ranging from 50.38× to 478.18×, demonstrating the potential of phase-change memristor-based in-memory computing for real-time neural modeling.