FALCON: A Fast and Low-Power Current-Mode Near-Sensor-Computing Architecture for Real-Time Edge Visual Processing

Near-sensor computing (NSC) has emerged as a promising paradigm for edge visual processing and data compression, to mitigate data transmission and computing overheads at IoT nodes. However, existing NSC still suffers from limited precision, reduced frame rate and low energy efficiency under complex DNN tasks due to inefficient analog memory, exponential computation overheads and considerable ADC burden. This paper introduces FALCON, a novel current-mode (CM) NSC architecture featuring in-current-register-processing (ICRP) unit and two-step multiply-and-accumulate (TS-MAC) for high-precision and low-latency feature extraction. Additionally, a reconfigurable ADC with embedded ReLU and pooling functionality is employed to improve ADC overhead and compression ratio. Implemented under a 55nm CIS process, FALCON achieves 12.92 TOPS/W with 7-bit weight precision and supports a frame rate of 3096 fps under 8 filters, with an iFOM of 10.1 pJ/pix•fps.