PERFORMANCE BENCHMARKING OF EMBEDDED EDGE DEVICES FOR VARIOUS FACE RECOGNITION MODELS

Abstract

Artificial Intelligence (AI) models are increasingly pivotal in enabling face recognition across various fields, from educational and research settings to public spaces.Effective deployment of these models requires high-performance hardware, such as RTX graphics cards or embedded edge devices like Nvidia's AGX Orin and Jetson Nano.This paper presents a comprehensive benchmarking study comparing the performance of these two devices, representing high and low-power edge computing options, using two face recognition models: ResNet and MobileNet.The benchmarking process assesses each model across two different input sizes deployed on both devices with varied configurations, including CPU thread allocation and GPU power distribution within containerized environments.Performance metrics such as inference time, GPU utilization, memory usage, and CPU load are analyzed to determine each device's suitability and efficiency.Additionally, model-specific parameters, including FLOPS, parameter count, and memory footprint, are examined to provide for an in-depth comparison.This paper presents detailed results and analyses of these performance indicators.