Deploying deep learning-based object detectors on low-compute edge AI SoCs remains challenging, as real-world performance depends on factors beyond nominal TOPS ratings, including architectural design, memory bandwidth, and system-level contention. This study presents a comprehensive and reproducible benchmarking of nine YOLO variants across three widely used Rockchip SoCs, covering multiple input resolutions, compute configurations, and operating conditions. Our results show that inference latency correlates more strongly with detection accuracy (mAP) than with FLOPs or parameter count, revealing the execution overhead introduced by recent architectural modules. Latency scaling with input size deviates from quadratic theoretical predictions due to bandwidth limitations, and multi-core NPU scheduling provides only marginal gains because of synchronization and shared-memory bottlenecks. Under multitasking stress, memory bandwidth emerges as the primary factor governing robustness, while energy-per-inference measurements highlight substantial efficiency differences across SoCs. These findings offer practical guidance for selecting and deploying object detection models on embedded platforms, emphasizing the need for hardware-aware model choices and memory-efficient optimizations in real-time edge AI applications.
Kong et al. (Wed,) studied this question.