Energy-Efficient Acceleration of Deep Learning based Facial Recognition on RISC-V Processor

Feature extraction at battery-powered resource-constrained end nodes is always challenging. Direct hardware mapping of a deep neural network (DNN) on an embedded platform faces difficulties in terms of computational power and memory. Hence, this work targets to accelerate deep learning MobileNet-based face recognition (FR) on RISC-V to optimize energy consumption by reducing execution time. To implement this, the Raspberry Pi-based FR system is developed to analyze the effect on recognition time and power consumption for an embedded platform compared to the CPU, for the same DL model. Recognition time increases 5.76× on Raspberry Pi compared to CPU. Hence, DNN algorithms require higher computation capabilities to reduce delay on resource-constrained embedded platforms. We have utilized a Sipeed Maixduino RISC-V artificial intelligence (AI) and Internet of Things (IoT) board. The Maixduino board can accelerate deep learning operations with the knowledge processing unit (KPU) accelerator. It was analyzed that the MobileNet-v1 0.75 on the KPU-based Maixduino RISC-V board decreases the classification time and power consumption by 30 ×and 2.86 ×, respectively, compared to CPU. Power-delay product (PDP) a.k.a. energy and energy-delay product (EDP) is an important parameter for the battery-operated IoT end nodes. For MobileNet on a RISC-V-based accelerator (Maixduino KPU), we achieved 82.75× and 2582.77× less PDP and EDP, respectively, compared to the CPU. Maixduino-based FR system with MobileNet is high-speed and energy-efficient in contrast to the current state-of-the-art (SoA)'s lightest (MobileNet-v1 0.25) model on the accelerator.