Comparison of edge computing platforms for hardware acceleration of AI: Kria KV260, Jetson Nano and RTX 3060

Abstract

As edge computing platforms become more extense and newer companies join the field, it becomes harder to know which platform to use in any specific case. These systems are often packed with a broad array of different computation architectures and different hardware acceleration technologies, this can be confusing at the moment of the election to integrate them as hardware accelerators in larger designs. Due to the efficiency of these platforms, they often enable creative problem-solving approaches to robotics and other fields where computation on the edge was not common that long ago. This thesis delves into leading hardware accelerators, analyzing the performance and power usage of three platforms: Kria KV260, Jetson Nano and RTX 3060. Experiments were conducted with two neural network models-ResNet-50 and YOLO-trained for image identification tasks. Our findings highlight the FPGA-based platform’s superior efficiency in terms of inference speed per watt compared to the other platforms.