FerroNDS Cuts Energy 25x Over SRAM With Analog Neural Dynamics

1.64 microjoules per neuron per inference at 200 Hz is the number that makes analog computing interesting again.

That's the per-neuron energy of FerroNDS, a 128-neuron neuromorphic system built on multi-bit ferrodiodes. It forecasts periodic, quasi-periodic, and chaotic signals over a 500 ms horizon while pulling sub-watt total power. Digital SRAM-based systems doing the same job need 25 to 40 times more silicon area.

Why Analog Wins for Time-Series

Neural dynamical systems are expressive temporal predictors, but their fine-grained state updates are a terrible fit for digital hardware optimized for dense matrix ops. Every time step means shuttling data between memory and compute. FerroNDS sidesteps that by using two analog primitives: an integrator for temporal accumulation and an oscillator for frequency-selective filtering. Both run in continuous time, matching the physics of the dynamical system itself.

The team mapped these primitives onto compute-in-memory hardware using multi-bit ferrodiodes. No SRAM banks, no data movement bottleneck. The analog domain handles the sequential structure natively.

FerroNDS Architecture: Integrator Plus Oscillator

A 128-neuron instance computes a short-time Fourier transform and then forecasts. At 200 Hz operating frequency, per-layer latency is 3.18 ms. Crank it to 10 kHz and latency drops to 63.87 µs, with per-neuron energy falling to 0.29 µJ. That's a 5.7x energy reduction just from higher throughput.

The real kicker: this is the first end-to-end integration of a ferrodiode into a neuromorphic computational framework. Prior ferroelectric work focused on memory elements or standalone devices. FerroNDS treats the whole thing as a continuous-time dynamical system, not a digital accelerator pretending to be analog.

Real-World Forecasting at Sub-Watt

Sub-watt real-time operation isn't theoretical. The paper demonstrates forecasting for periodic, quasi-periodic, and chaotic signals. Chaotic prediction is the hardest case, and FerroNDS handles it at the same energy budget. That matters for edge deployments where power is capped and you can't afford a GPU.

Area reduction alone makes this interesting. SRAM-based digital alternatives need 25-40x more die area for the same neuron count. Analog compute-in-memory with ferrodiodes packs more compute into a smaller footprint, which translates directly to lower cost and higher density for sensor nodes or wearable devices.

What This Enables Next

FerroNDS proves ferroelectric compute-in-memory is a practical substrate for analog neural dynamical systems. The next step is scaling beyond 128 neurons and integrating the system onto a single die with the sensor frontend. That would put real-time chaotic forecasting into power-constrained systems like drone flight controllers or implantable medical monitors.

Source: Neural dynamical systems on ferroelectric compute-in-memory for real-time forecasting
Domain: arxiv.org