A 94.8nW Battery-Free Intelligent Silicon Platform Enabling Distributed, Adaptive, and Event-Driven Multimodal Sensing at the Edge
Haochen Zhang, Wei-Han Yu, Zhongyu Zhao, Zhizhan Yang, Ka-Fai Un, Jun Yin, Rui Paulo Martins, Pui‐In Mak · 2025
Sensor nodes with machine learning (ML) are adept at analyzing intricate environmental and physiological data patterns at the edge [1]–[9]. The design of such ultra-low-power (ULP) devices strives to reduce power consumption, which ensures continuous and energy-harvested operation even with fluctuating ambient available energy levels down to 100nW [10]–[11]. Consequently, ML capabilities on such ULP devices are constrained to perform lightweight detection for events such as voice activity [3] [6], arrhythmia [4] [8], and bearing anomalies [9]. Yet, these isolated, monomodal sensing paradigms suffer from low task complexity and accuracy for overlooking fused information from spatially distributed sensor networks. Additionally, real-world applications subject to data distribution drifts necessitate model adaptability to maintain accuracy in ever-changing environments [12]. Also, current ULP neural network (NN) accelerators consume significant power, limiting their ability to expand the network size, while impeding the overall ML performance.