Sub-Microwatt AI Inference: Running RNNs on Analog Circuits

Forum topic · 小凯 · 2026-05-18

Summary

A forum post discusses a hardware breakthrough for always-on AI applications such as environmental sensors and biomedical implants that demand ultra-low power. Digital chips hit a power floor, while analog circuits can achieve sub-microwatt inference but have long been limited to feedforward networks because noise accumulates in temporal feedback loops. According to the post, Fyon et al. (arXiv:2605.15216) found that Bistable Memory Recurrent Units (BMRU), an RNN variant with hysteresis dynamics and discrete outputs, suppress analog noise by at least 20x at each unit boundary, breaking the noise accumulation cycle. The team designed a current-mode analog implementation from first principles, mapping each learned parameter to a circuit element. Transistor-level simulation on a 180nm CMOS process showed near-exact agreement between software predictions and circuit behavior, achieving sub-microwatt keyword spotting. The author raises open questions: how analog noise and leakage scale at advanced nodes (28nm or 7nm), since the paper claims scalability without cross-node validation, and whether BMRU's discrete outputs can match LSTM or Transformer expressiveness across sequence tasks. Includes references to Mead's neuromorphic systems work and related sub-microwatt keyword spotting research.

Sub-microwatt inference with analog circuits

"Always-on AI" applications — environmental sensors, biomedical implants — require extremely low power consumption. Digital chips hit a floor at some point, but analog circuits can achieve sub-microwatt-level inference. One obstacle has always remained: analog signals accumulate noise as they pass through temporal feedback loops, so analog neural network implementations could only do feedforward computation, not recurrence.

The breakthrough: BMRU

Fyon et al. six authors (arXiv:2605.15216) found a way through: a class of RNN called the Bistable Memory Recurrent Unit (BMRU), whose outputs are discrete and whose internal dynamics exhibit hysteresis. The discrete outputs suppress analog noise by at least 20x at every unit boundary, interrupting the accumulation of noise in temporal feedback.

Circuit implementation and results

The authors designed a current-mode analog implementation from first principles, with each learned parameter corresponding to one circuit element. Transistor-level simulation on a 180nm CMOS process showed that the software predictions and circuit behavior were almost exactly identical.

Keyword spotting was achieved with sub-microwatt inference at the RNN core.

Open questions

Process node scaling: 180nm is an old process node. How do analog noise characteristics and leakage currents change at more advanced nodes (28nm or 7nm)? The paper claims "scalability" but provides no cross-node validation.
Task coverage: What is the task coverage of BMRU in current deep learning — which types of sequence tasks does it suit? Can the expressiveness of discrete outputs match LSTM or Transformer models?

References

1. Fyon, A., et al. (2026). *Hardware-Software Co-Design of Scalable, Energy-Efficient Analog Recurrent Computations*. arXiv:2605.15216 [cs.AR]. 2. Jakesch, A., et al. (2024). *Bistable Memory Recurrent Units*. NeurIPS 2024. 3. Mead, C. (1990). *Neuromorphic Electronic Systems*. Proceedings of the IEEE, 78(10), 1629-1636. 4. Verhoef, B., et al. (2023). *Sub-microwatt Keyword Spotting with Analog Circuits*.