Neural dynamical systems on ferroelectric compute-in-memory for real-time forecasting

2026-06-15

AI Summary

This work addresses the inefficiency of conventional digital hardware in supporting the continuous-time state evolution required by neural dynamical systems, which limits energy efficiency and latency performance for real-time signal prediction. The authors propose FerroNDS—an analog neuromorphic system based on ferroelectric compute-in-memory architecture—that, for the first time, integrates ferroelectric diodes end-to-end into a neural dynamical framework. By leveraging analog integrator and oscillator primitives, FerroNDS enables continuous-time dynamic modeling. The system employs multi-bit ferroelectric diodes to realize an in-memory computing architecture with 128 neurons, capable of predicting periodic, quasi-periodic, and chaotic signals up to 500 ms in real time. Operating at 10 kHz, it achieves per-neuron inference energy as low as 0.29 μJ, reduces area by 25–40× compared to SRAM-based digital counterparts, and attains a single-layer latency of merely 63.87 μs.

Abstract

Neural dynamical systems are expressive temporal predictors that capture continuous-time dynamics through fine-grained state updates. However, this sequential structure maps poorly onto digital hardware optimized for dense matrix operations, a mismatch that analog neuromorphic computing, with its native continuous-time dynamics, can resolve. We introduce FerroNDS, a neuromorphic system built from two analog primitives: an integrator for temporal accumulation and an oscillator for frequency-selective filtering. We map this system onto compute-in-memory hardware based on multi-bit ferrodiodes. A 128-neuron instance of FerroNDS computes short-time Fourier transform and forecasts a 500-ms horizon for periodic, quasi-periodic, and chaotic signals. The system achieves sub-watt real-time operation with per-neuron per-inference energy of 1.64 $μ$J (200 Hz) and 0.29 $μ$J (10 kHz), 25-40$\times$ area reduction over SRAM-based digital systems, and per-layer latency of 3.18 ms (200 Hz) and 63.87 $μ$s (10 kHz). To our knowledge, this is the first end-to-end integration of a ferrodiode into a neuromorphic computational framework, establishing ferroelectric compute-in-memory as a practical substrate for analog neural dynamical systems.

Problem

• neural dynamical systems
• real-time forecasting
• compute-in-memory
• analog neuromorphic computing
• ferroelectric devices

Innovation

• ferroelectric compute-in-memory
• neural dynamical systems
• analog neuromorphic computing
• ferrodiode
• real-time forecasting