Multi-GPU Benchmarking of YOLO Generations for PCB Defect Detection

Thorough defect detection on printed circuit boards (PCBs) is required to ensure manufacturing quality and reliability. Even though many modern YOLO-based detectors exist, very little research focuses on comparative evaluation across multiple YOLO generations while accounting for deployment-related factors such as training costs, hardware limitations, and energy efficiency. Twelve model configurations from YOLOv8, YOLO11, YOLOv12, and YOLO26 generations have been benchmarked in this study on a dataset of 10,628 images representing six types of PCB defects. All the configurations have been trained under similar conditions, including transfer learning and data augmentation. Metrics utilized in the performance assessment include Precision, Recall, F1-score, [email protected], and [email protected]:0.95. Deployment analysis considers inference speed, training time, VRAM usage, GPU speedup, and energy consumption for NVIDIA RTX 3060, RTX 5060 Ti, and RTX 4090 GPUs. All models achieved excellent detection performance, with [email protected] ranging from 0.985 to 0.989. YOLOv8m obtained the highest [email protected]:0.95 value of 0.697. RTX 5060 Ti demonstrated the best balance between training speed and energy efficiency, while RTX 4090 showed the highest acceleration for complex models. The current research is the first to provide multi-GPU and multi-YOLO benchmarking for real-life PCB defect detection in industrial PCB inspection.