This Chip Reconstructs the Brain in Under 10 Milliseconds

Imagine watching a brain fold and unfold on a screen in less time than it takes to blink. That’s the scale these researchers are hitting: a memristor-based chip that models the brain’s convoluted cortex in under 10 milliseconds.

The work, led by Yuchao Yang at Peking University with collaborators at the Shanghai Institute of Microsystem and Information Technology, was published in Science as "A sub-10-millisecond neural dynamical system based on phase-change memristors." Instead of shuttling data back and forth between memory and a separate processor—the traditional bottleneck in numerical modeling—the team pushes calculations into the memory itself, using phase-change memristors to store and compute simultaneously.

Why does that matter? Reconstructing the brain’s folded surface in real time is computationally brutal. Neural dynamical systems, which couple neural networks with differential equations to track evolving physical systems, require repeated integration steps, error checks, and finely tuned step sizes. Move data across buses for every step and you waste time and power. Do the math where the data lives, and everything speeds up.

The chip performs neural dynamics at millisecond timescales, matching the cadence of biological computation.

On silicon, the team used a 40-nanometer process to pack in-memory computing and conductance-drift memristor arrays into just 0.28 square millimeters. The device runs at 50 MHz and completes each numerical integration through nine pipeline stages. The payoff was dramatic: during neural dynamics workloads the hardware ran between about 3.8× and 36.3× faster than top-tier application-specific integrated circuits, while drawing roughly 11.8× to 24.7× less power. For the specific task of cortical surface reconstruction, the chip reported speedups as high as 478× compared with an NVIDIA A100 GPU.

Overview of NDS hardware with multilevel and fine-grained CCD memristor. 

Performance numbers are one thing. Quality is another. The researchers reconstructed the boundaries between white and gray matter and produced three-dimensional manifold-based surface meshes in real time. The meshes were smooth and closed, preserving the brain’s folds and maintaining topological consistency. Quantitative metrics—average symmetric surface distance and Hausdorff distance—showed the reconstructions closely matched target anatomies, indicating the approach doesn’t sacrifice fidelity for speed.

Think of the potential. Real-time cortical mapping could change how we approach surgical guidance, brain–computer interfaces, and live medical imaging. Digital brain twins that update as a surgery progresses or adaptive imaging tools that respond in milliseconds are suddenly plausible, not distant engineering dreams. For researchers studying neurodegenerative diseases like Alzheimer’s or Parkinson’s, faster, high-fidelity surface reconstruction could mean new ways to monitor subtle structural changes as they happen.

There are practical hurdles ahead: scaling the device for larger models, integrating with existing clinical workflows, and validating robustness across diverse patient anatomies. Yet the central idea—collapse the memory-processor divide and let the hardware do the heavy math where the data lives—feels less like incremental improvement and more like a shift in how we design neurocomputational tools.

If machines can now mimic the brain’s timing while preserving anatomical detail, what new interactions between clinicians, patients, and algorithms will emerge in the next decade?