A sub-10ms neural dynamical system based on phase change memristors

A sub-10ms neural dynamical system based on phase change memristors Introduction High-fidelity geometry under strict topological constraints remains a critical challenge for realizing physical-world models, particularly for accurate genus-0 homeomorphic deformations that require real-time, dense, and smoothly differentiable deformation fields on manifolds. Neural dynamical systems (NDS) excel at these tasks using adaptive-stepsize numerical integration with embedded neural networks (ENNs), but causing heavy data movements and iterative stepsize searches. As a result, the NDS runtime remains on the order of hundreds of milliseconds with classical ENNs based on multi-layer perceptrons, limiting practical deployments. Here we report a low-latency NDS of sub-10ms leveraging two key attributes of phase-change memristors (PCM): conductance drift and multilevel conductance.