Edge AI Inference and Industrial Vision
SimtoLink AI edge inference hardware uses a dedicated NPU (neural processing unit) for inference at the edge. Single-frame latency is under 50 ms, and the models include YOLO, ResNet, and MobileNet. Images stay on site and are not uploaded to the cloud, which matches manufacturing requirements for latency, data compliance, and privacy.
Several video streams can be inferred in parallel, and the compute rating is chosen for the stream count. These devices are already in electronics manufacturing, food packaging, textile dyeing and printing, and package printing: more than 5,000 units deployed, and more than 100 million images a day. Inference performance depends on model complexity and input resolution. SimtoLink compresses and quantizes models so a limited compute budget still reaches the inference performance you need.
Model Frameworks and How They Are Deployed
Models exported from TensorFlow, PyTorch, and Caffe are converted through ONNX (Open Neural Network Exchange) and then loaded on the device. NVIDIA TensorRT optimization, including INT8 quantization, raises inference speed by 2～3 times and cuts the compute you need while accuracy holds.
Common industrial defect models and OCR models are preinstalled and ready to run. Hot Swap model updates change the version without a shutdown. Several models can be loaded together, so object detection, image classification, and instance segmentation can run at once. SimtoLink helps with migration, pruning, and distillation, and a new framework can be added in a container.
Typical Industry Uses
The same hardware covers SMT lines that want full solder-joint inspection instead of sampled AOI, food-pack inspection, and retail shelf checks. The cases below are typical uses and the results after deployment:
PCB Solder Joints on an SMT Line
Full inspection of PCB solder joints replaces sampled AOI. One electronics plant put in 20 AI edge devices and cut the miss rate from 2% to 0.05%, which tightened yield control.
Food-Pack Label Inspection
Wrong, missing, and skewed labels are caught as the pack moves. One food company used the devices on packaging and cut defective packs leaving the plant by 80%.
Retail Shelf Checks
Planogram compliance, out-of-stock alerts, and SKU recognition. Images are processed on site and only the summary is uploaded, which keeps privacy and cuts bandwidth.
Textile Dyeing and Printing Defects
Surface flaws, shade differences, and stains on cloth are detected for live sorting and quality traceability. The installation has run at the customer site for more than 12 months.
Choosing Compute
Match compute to model size, input resolution, and the frame rate you need. The table is a starting point for typical jobs:
| Application | Suggested compute | Main factors |
|---|---|---|
| One 1080p stream at 30 fps, object detection | About 4 TOPS | Light model, real-time |
| Several streams, or a higher resolution | 12 TOPS and above | Parallel streams, higher-accuracy models |
| Complex defect detection or instance segmentation | 16+ TOPS | Large models, high-resolution input |
| Light OCR or classification | 2～4 TOPS | Small models, infrequent inference |
Validate the job on a lightweight model first, then size the device to the frame rate and accuracy you need. Custom models can be deployed with Docker when you need to extend later. SimtoLink offers a compute-sizing aid and a POC, and for harder cases model optimization and custom training.
Capabilities and What Deployment Needs
- Low-latency inference: Inference stays at the edge, with single-frame latency <50ms, for live quality checks and closed-loop control on the line.
- Data compliance: Images do not leave the plant, which meets data-localization, privacy, and industry-regulation requirements.
- Hot model updates: Remote OTA updates and a staged rollout cut risk, so a new model version does not need a site visit.
- Parallel streams: Several video streams are inferred together, and the compute rating is chosen for the stream count.
- Ongoing tuning: Accuracy is reviewed on a schedule, failure cases are collected for the next model, and SimtoLink can train and tune to order.
How a Project Usually Moves
A practical sequence for an AI edge inference project:
For a new line, pilot a small batch and scale only after the result holds. Choose the device from the model architecture, the frame rate, and the accuracy you need. SimtoLink can guide training, suggest how to label data, and bring an industry solution, so the path from proof to volume production stays short.
Services
SimtoLink helps select the AI edge device, deploy the model, and tune it, including training, INT8 quantization, a POC, and schemes built for an industry. Product manuals, a compute-sizing aid, and a scheme quote are available.