Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces FerroNDS, a neuromorphic system that leverages analog computing with ferroelectric compute-in-memory hardware to effectively model neural dynamical systems for real-time forecasting. By utilizing an integrator and an oscillator, FerroNDS achieves efficient temporal accumulation and frequency-selective filtering, enabling it to forecast signals over a 500-ms horizon with significantly reduced energy consumption and latency compared to traditional digital systems. The results show a remarkable area reduction of 25-40 times over SRAM-based architectures, marking a significant advancement in the practical application of analog neuromorphic computing for dynamic signal prediction.
FerroNDS achieves real-time forecasting with over 25 times less area than conventional digital systems while operating at sub-watt power levels.
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.