Energy-Efficient Edge Inference on Multi-Channel Streaming Data in 28nm HKMG FeFET Technology

Abstract

We present a system implementing extremely energy-efficient inference on multi-channel biomedical-sensor data. We leverage Ferroelectric FET (FeFET) to perform classification directly on analog sensor signals. We demonstrate: (i) voltage-controlled multi-domain ferroelectric polarization switching to obtain 8 distinct transconductance $(\text{g}_{\text{m}})$ states in a 28nm HKMG FeFET technology [1], (ii) 30x tunable range in $\text{g}_{\text{m}}$ over the bandwidth of interest, (iii) successful implementation of artifact removal, feature extraction and classification for seizure detection from CHB-MIT EEG dataset with 98.46% accuracy and $< 0.375/\text{hr}$. false alarm rate for two patients, (iv) ultra-low energy of 47 fJ/MAC with 1,000x improvement in area compared to alternative mixed-signal MAC.