[en] We present a hardware–algorithm co-design approach for ultra-low-power analog recurrent neural networks, leveraging First Quadrant Bistable Memory Recurrent Units (FQ-BMRUs) whose discrete and hysteretic dynamics map efficiently onto subthreshold current-mode circuits. Implemented in 180 nm CMOS, the proposed system achieves over 90% accuracy on keyword spotting at ~100 nW, demonstrating robust noise suppression and orders-of-magnitude energy savings compared to conventional digital approaches.
Brandoit, Julien ✱; Université de Liège - ULiège > Montefiore Institute of Electrical Engineering and Computer Science
Fyon, Arthur ✱; Université de Liège - ULiège > Montefiore Institute of Electrical Engineering and Computer Science
Mendolia, Loris ; Université de Liège - ULiège > Montefiore Institute of Electrical Engineering and Computer Science
Ernst, Damien ; Université de Liège - ULiège > Département d'électricité, électronique et informatique (Institut Montefiore) > Smart grids
Redouté, Jean-Michel ; Université de Liège - ULiège > Département d'électricité, électronique et informatique (Institut Montefiore) > Systèmes microélectroniques intégrés
Drion, Guillaume ; Université de Liège - ULiège > Département d'électricité, électronique et informatique (Institut Montefiore) > Systèmes et modélisation
✱ These authors have contributed equally to this work.
Language :
English
Title :
Hardware-Algorithm Co-Design of Analog Recurrent Neural Networks for Sub-Microwatt Inference
Publication date :
June 2026
Number of pages :
1
Event name :
International Conference on Neuromorphic Computing and Engineering
F.R.S.-FNRS - Fonds de la Recherche Scientifique FPS BOSA - Federal Public Service Policy and Support FRIA - Fonds pour la Formation à la Recherche dans l'Industrie et dans l'Agriculture
JB, LM and AF are, respectively, a FRIA grantee, a FRIA grantee and a Research Fellow of the Fonds de la Recherche Scientifique – FNRS. This work was supported by the Belgian Government through the Federal Public Service Policy and Support grant NEMODEI.
Commentary :
This work has been the subject of a patent application under number EP26175243.0. (HARDWARE) and EP26175248.9. (SOFTWARE)
