PCB Defect Detection: Vision AI on GPU

The Challenge: 18,000 Boards, 200 Components Each

A contract electronics manufacturer in South Wales runs three SMT (surface-mount technology) production lines, producing 18,000 populated PCBs per day for automotive, medical, and industrial clients. Each board contains 150 to 350 components, and defects — solder bridges, cold joints, missing components, polarity reversals, tombstoned resistors — must be caught before boards leave the factory. The existing automated optical inspection (AOI) system, installed eight years ago, generates a false positive rate of 12%, meaning operators spend hours each shift verifying defects that do not exist. Worse, its detection sensitivity on fine-pitch QFN packages has degraded, and two client RMAs last quarter were traced to solder defects the AOI missed entirely. A new AOI system costs £250,000–£400,000 per line.

The manufacturer wants to layer AI-powered secondary inspection over their existing AOI — using the same camera images but running them through a modern deep learning model that reduces false positives while improving genuine defect capture rates.

AI Solution: GPU-Accelerated Visual Defect Detection

Object detection models such as YOLOv8, DETR, or custom-trained RetinaNet variants excel at PCB defect detection when trained on factory-specific imagery. The model processes high-resolution board images captured by existing AOI cameras, identifies defect regions, classifies defect types, and outputs bounding boxes with confidence scores. A vision model running on a dedicated GPU server positioned on the factory network processes each board image in under 100ms — fast enough to keep pace with the pick-and-place machine’s throughput.

The AI layer receives every image the AOI flags as defective and applies a second opinion. Genuine defects pass through; false positives are filtered out. Additionally, the AI model runs on a random sample of AOI-passed boards to catch defects the legacy system misses.

GPU Requirements

PCB inspection images are typically captured at high resolution (4096×3072 or higher) and may require processing at full resolution to detect sub-millimetre defects. YOLOv8-Large at full resolution demands significant GPU compute for real-time throughput.

Even at 18,000 boards daily, a single RTX 5090 has massive headroom. The spare capacity allows running multiple model variants — one optimised for solder defects, another for component placement — simultaneously. Private AI hosting ensures client-specific board designs and proprietary defect data remain within GDPR-compliant infrastructure.

Recommended Stack

• YOLOv8 or RT-DETR for real-time defect detection, trained on annotated factory imagery.
• Ultralytics framework for training, validation, and export to TensorRT for maximum inference speed.
• NVIDIA TensorRT for optimised inference, reducing latency by 40-60% compared to native PyTorch.
• Label Studio or CVAT for defect annotation, with factory operators labelling images during calibration.
• Optional: an LLM via vLLM for generating human-readable defect reports from detection results.

For traceability, integrate with document AI to read board serial numbers and lot codes from PCB silkscreen markings, linking defect data to specific production batches.

Cost Analysis

A new AOI system costs £250,000–£400,000 per line (£750,000–£1.2 million for three lines). An AI overlay using existing cameras and a dedicated GPU server costs a fraction of that, delivering equivalent or superior detection accuracy. The 12% false positive reduction alone saves approximately 3 hours of operator time per shift — roughly £45,000 annually across three shifts.

Preventing client RMAs delivers even larger savings. Each automotive RMA costs the manufacturer an average of £35,000 in investigation, replacement, and reputational damage. Catching two additional defect escapes per quarter saves £280,000 annually.

Getting Started

Begin by capturing 10,000 annotated board images from your existing AOI system — both defect images and known-good boards. Train a YOLOv8 model on your specific board designs and defect types. Run the AI in parallel with the existing AOI for 30 days, comparing detection rates and false positive rates. Most manufacturers achieve production-ready accuracy within two training iterations.

GigaGPU provides UK-based dedicated GPU servers configured for industrial vision workloads. Deploy on-premises or in a UK data centre with guaranteed latency, and add an AI chatbot for operator query support on the factory floor.