What is edge computing in OT?

Industrial edge software like N3uron runs on a wide range of hardware, which is one of its key advantages over fixed-function gateway appliances. For small sites with limited tag counts (fewer than 5,000 tags), N3uron runs comfortably on ARM-based single-board computers or small fanless industrial PCs with 1-2 GB RAM and 8-16 GB of storage. For large sites with tens of thousands of tags, high-frequency data, or local analytics workloads, a standard x86 industrial PC or server-class hardware is appropriate.

The hardware should be rated for the operating environment: DIN-rail mounted fanless PCs for control panel installation, ruggedized IPCs for harsh factory floors, and Class 1 Division 2 certified devices for hazardous locations in oil and gas. SWTB's hardware partners include suppliers whose equipment has been validated against N3uron and TOP Server for specific industrial environments.

Store-and-forward is a buffering mechanism that preserves data collected at the edge during periods when the upstream connection is unavailable. When the edge node loses connectivity to its upstream destination (MQTT broker, historian, cloud platform), it continues collecting data from field devices and writing it to a local circular buffer with original source timestamps. When connectivity is restored, the edge node delivers the buffered data in chronological order, typically flagging it as historical data so the receiving system stores it at the correct timestamp.

The buffer size is configurable and typically sized based on the expected maximum outage duration multiplied by the average data rate. N3uron's store-and-forward buffer persists across power cycles, so an unplanned plant shutdown does not lose the buffered data. Cogent DataHub provides an additional store-and-forward layer at the DMZ crossing point for architectures where data must survive both the edge-to-DMZ link outage and the DMZ-to-enterprise link outage independently.

A traditional OPC server (like TOP Server) focuses on device connectivity: connecting to field devices using their native protocols and exposing the data as OPC tags for OPC clients to poll. It is primarily a connectivity layer, not a processing layer.

An edge platform (like N3uron) does everything an OPC server does and adds the processing pipeline on top: deadband filtering and report-by-exception, ISA-95 data modeling and contextualization, store-and-forward buffering, MQTT and Sparkplug B publishing, local historian, and scripting for derived metrics and logic. In practice, SWTB customers often run both: TOP Server or Kepware Edge as the device connectivity layer, and N3uron as the edge processing layer consuming the OPC data from TOP Server, contextualizing it, and publishing it upstream.

Yes, and this is one of the fastest-growing edge computing use cases in OT environments. The typical pattern is: train an anomaly detection or predictive maintenance model in the cloud using historical data, export it in a standard format (ONNX, TensorFlow Lite), and deploy it to the edge node for inference. The edge node runs the model on live sensor data continuously, generating predictions and alerts in near-real time without a cloud round trip.

N3uron's scripting engine can execute Python scripts that load and run ML models on incoming tag data. For more intensive inference workloads, edge hardware with GPU or NPU acceleration is increasingly available at industrial-grade form factors. The model itself is updated from the cloud when a new version is trained, while inference continues to run locally between updates.

N3uron's published benchmarks show it handling 50,000 tags on modest hardware (an Intel NUC-class device) at 1-second scan rates without saturating CPU or memory. On more capable hardware, tag counts in the hundreds of thousands are achievable. The limiting factor in practice is usually not the software but the number of concurrent device connections and the network capacity on the plant LAN.

For very large sites, it is common to deploy multiple edge nodes, each responsible for a specific area or production line, and aggregate them at a site-level N3uron instance or MQTT broker. N3uron's architecture supports multi-tier deployments natively, with edge nodes at the area level publishing to a site-level broker that in turn feeds the enterprise UNS.

Complementary, not a replacement. Edge and cloud serve different roles in industrial data architecture and are most effective when designed to work together. The edge handles time-sensitive, local, and offline-resilient processing: data collection, filtering, contextualization, buffering, local alarming, and fast-loop inference. The cloud handles long-term storage, cross-site analytics, AI model training on large historical datasets, enterprise reporting, and functions that require a global view of all sites simultaneously.

The practical architecture is an edge node at each site or area that sends filtered, contextualized data to a cloud data platform. The cloud receives a fraction of the raw data volume (because of deadband filtering and report-by-exception), but the data it receives is already structured and semantically meaningful, rather than raw device addresses and counts.