Audio and Image Cross-Modal Intelligence via a 10TOPS/W 22nm SoC with Back-Propagation and Dynamic Power Gating

Zichen Fan, Hyochan An, Qirui Zhang, Boxun Xu, Li Xu, Chien-Wei Tseng, Yimai Peng, Ang Cao, Bowen Liu, Changwoo Lee, Zhehong Wang, Fanghao Liu, Guanru Wang, Shenghao Jiang, Hun-Seok Kim, David T. Blaauw, Dennis Sylvester

Abstract: We present an ultra-low-power multimedia signal processor (MMSP) SoC that integrates a versatile deep neural network (DNN) engine with audio and image signal processing accelerators for cross-modal IoT intelligence. The proposed MMSP features 2MB MRAM to store all DNN weights on-chip with an energy-efficient dataflow using an MRAM-cache and dynamic power gating. The SoC achieves up to 3-10 TOPS/W peak energy efficiency and consumes only 0.25-3.84 mW. Being the first to demonstrate CNN, GAN, and back-propagation (BP) on a single accelerator SoC for cross-modal fusion, it outperforms state-of-the-art DNN processors by 1.4 - 4.5× in energy efficiency.