Exploring Power-Performance Trade-offs in Softcore RISC-V with CNN Accelerators for Digital Agriculture and Smart Farming

Abstract

This study presents a detailed exploration of RISC-V softcore assisted by Convolutional Neural Networks (CNNs) accelerators for image classification in tasks associated with digital agriculture, including land cover classification, crop disease detection, and seed quality analysis. Like other sectors, agriculture is becoming a knowledge-intensive enterprise. We address the computational and energy demands of CNNs where processing is offloaded to the edge, often using battery-powered systems embedded in sensors, field equipment, and drones. By integrating the RVX RISC-V softcore with FINN-generated accelerators on an Artix-7 FPGA, we evaluate the trade-offs between resource utilization, throughput, and power e efficiency. Our results demonstrate that dedicated accelerators operating in high-speed burst modes can improve power efficiency (FPS/W) by up to 57% in relatively complex agricultural CNNs, although at a steep cost in FPGA resources, and that the integration with a softcore processor provides the necessary flexibility for data management and system control, enabling a versatile platform for Smart Farming applications.