Evaluating the Efficiency of Hailo-8 for YOLOv10-Based Object Detection in Edge Environments

Abstract

This work will examine the real-time performance of object
detection based on the Hailo-8 AI accelerators, a low-power
inference developed for edge computing. The system performance
was tested in terms of system throughput, system
latency, and CPU load performance metrics in an Intel Core
i9-10900K system running Ubuntu 22.04. On a YOLOv10
model with 80 classes of objects in the COCO dataset, an
average of 12.48 FPS with a total system latency of 57.39
ms per frame was recorded, with 56.40 ms spent in inference
and 0.98 ms for CPU processing. This clearly indicates that
Hailo-8 significantly outperforms CPU-only inference with
low power consumption performance.