# Manufacturing industrial IoT architecture decision matrix

Pick **IoT SiteWise**, **custom lake**, or **IoT TwinMaker** placement for plant-floor
telemetry by tag count, latency needs, and OEE maturity.

> Reflects **July 2026**: IoT SiteWise **multivariate anomaly detection** GA (Jul 2025) in
> us-east-1, eu-west-1, ap-southeast-2; up to **300 sensors per model**; training minimum
> 1 hour; Greengrass v2 SiteWise gateway deployment path.

## 1. Primary pain

| Pain | Default stack | Escalate when |
|------|---------------|---------------|
| OEE / downtime visibility | **IoT SiteWise** + QuickSight | < 100 tags — Athena on S3 batch may suffice |
| Predictive maintenance | SiteWise **anomaly detection** | No 90-day clean history — fix tagging first |
| Edge buffering during outage | **Greengrass v2** + SiteWise gateway | Stable plant LAN — cloud-only ingest OK |
| Digital twin visualization | **IoT TwinMaker** | No 3D asset model budget — defer TwinMaker |
| OT security segmentation | **IoT Greengrass** + dedicated VPC | IT-only VLAN flat — fix network before cloud |

**Opinionated default:** **SiteWise for asset hierarchy + OEE**, **Greengrass at the edge**
for OPC-UA buffering, **defer TwinMaker** until OEE baselines exist.

## 2. When NOT to deploy full SiteWise

| Situation | Build lighter |
|-----------|---------------|
| < 50 OPC-UA tags, one line | IoT Core → Timestream → Grafana |
| No asset model discipline | Fix tag naming before SiteWise model import |
| Partner forbids cloud egress | Greengrass local historian only — no cloud path |
| Logistics / fleet tracking | Use [logistics architecture](/blog/aws-logistics-supply-chain-reference-architecture-2026/) |

## 3. Integration pattern

| Source | Edge | Cloud | Latency | Pitfall |
|--------|------|-------|---------|---------|
| OPC-UA (PLC) | SiteWise Gateway on Greengrass | IoT SiteWise | Seconds–minutes | Cert rotation at scale |
| MQTT (legacy) | IoT Core rules | SiteWise stream | Seconds | Topic sprawl |
| MES batch export | S3 landing | Glue → Athena | Hours | Silent extract failure |
| Manual OEE sheets | QuickSight dataset | — | Days | Do not skip digitization |

## 4. Anomaly detection fit

SiteWise native anomaly detection ([Jul 2025 GA](https://aws.amazon.com/about-aws/whats-new/2025/07/aws-iot-sitewise-multivariate-anomaly-detection/)):

- **Multivariate** models across up to **300 sensors**
- Training billed minimum **1 hour**, prorated per minute
- High-frequency inferencing schedules (5–60 min) vs low-frequency count-based

Use when: rotating equipment (motors, compressors) with continuous telemetry history.
Skip when: < 30 days labeled data or event-based failures only (use EventBridge rules instead).

## Related posts

- [Manufacturing IoT reference architecture 2026](/blog/aws-manufacturing-industrial-iot-reference-architecture-2026/)
- [SiteWise anomaly detection](/blog/aws-iot-sitewise-native-anomaly-detection-predictive-maintenance/)
- [OT/IT convergence](/blog/ot-it-convergence-aws-architecture-patterns/)
- [Manufacturing industry hub](/industries/aws-manufacturing/)