Why compute moves toward the line
“Edge” simply means running software close to where the data is produced, on the plant network rather than in a distant data center. Four reasons usually drive the decision:
- Latency. A decision that must happen in milliseconds cannot wait for a round trip to a remote platform.
- Resilience. A line should keep running and keep recording when the wide-area link is down.
- Bandwidth. Filtering, aggregating and compressing near the source means less data to send. The IIoT data budget calculator shows how much the sampling choices matter.
- Locality. Industrial protocols live on the plant network, and some data should stay on site.
What belongs at the edge
- Protocol conversion, for example Modbus or OPC UA on one side and MQTT on the other.
- Buffering and store-and-forward, so an outage does not become a data gap.
- Filtering, deadband and local aggregation.
- Local alerts and simple visualization for the operators next to the machine.
- Inference with a trained model where latency or connectivity demands it.
Model training, cross-plant analytics and long-term storage normally stay central, where compute and storage are easier to scale.
Choosing hardware
| Class | Typical use | Checks before buying |
|---|---|---|
| DIN-rail gateway | Protocol conversion, buffering for a cell | Serial ports for RS-485, two Ethernet ports for segmentation, 24 VDC input |
| Fanless industrial PC | Containers, local databases, dashboards | Operating temperature, storage endurance, memory, remote management |
| GPU-capable module | Vision or other model inference | Power and cooling budget, driver and OS support lifetime |
Whatever the class, check the operating temperature range, absence of moving parts, mounting, power input protection, certifications your site requires and how long the vendor will keep the product available and patched.
Software on the edge
- Run workloads in containers with explicit CPU and memory limits, so one fault cannot starve the others (see containerization on the factory floor).
- Use a minimal, hardened operating system, with unused services off.
- Synchronize the clock (time synchronization for industrial data) and rotate local logs so a full disk never stops the gateway.
Getting data out
Prefer outbound-only connections from the edge to a central broker, secured with TLS and per-device certificates, with local buffering for outages. Agree on topic and payload conventions up front: MQTT topic design for factories covers a structure that scales.
Operating it is the real cost
Every edge node is an asset that must be patched, monitored and replaced. Plan the update pipeline (CI/CD for edge devices), monitor heartbeat, disk and memory, keep configuration in version control and write down how to swap a failed unit and restore its configuration.
When not to use edge
If latency and connectivity are not constraints and a central platform is easy to reach, adding edge nodes just adds things to maintain. Weigh the options with the edge versus cloud decision framework.
A starter checklist
- Name the workload and the reason it must run near the line.
- Choose hardware from the environment, not from the data sheet headline.
- Define resource limits, update method and rollback before installation.
- Use outbound-only, encrypted connections with local buffering.
- Assign an owner for monitoring and replacement.