A Heterogeneous TinyML SoC with Energy-Event-Performance-Aware Management and Compute-in-Memory Two-Stage Event-Driven Wakeup

Yanchi Dong, Xueping Liu, Kangbo Bai, Guoxiang Li, Meng Wu, Yiqi Jing, Yihan Zhang, Pixian Zhan, Yadong Zhang, Yufei Ma, Ru Huang, Le Ye, Tianyu Jia

Abstract: This paper presents a heterogeneous TinyML SoC with E2p- aware system-level energy management achieving a minimum 3.5µW and 30000× peak-to-idle power ratio. The energy-event-performance (E2P)-aware management utilizes various fully synthesizable monitors to be aware of the runtime status of heterogeneous blocks and hierarchical voltage regulation for MEP search at system level, which gains >28% energy saving compared to single block MEP. A 2-stage CIM-based event-driven wakeup scheme is also developed to reduce the always-on energy by over 87%. The presented TinyML SoC is suitable for edge AI applications with state-of-the-art low-power features.