Edge AI Reference Architecture: 8-Camera PoE Deployment

Last updated: February 2026

TL;DR

An 8-camera edge AI node requires careful co-design of compute, networking, power, and storage. At this scale, the bottlenecks shift from single-camera inference performance to network bandwidth, PoE power budget, thermal envelope, and storage write throughput. This reference architecture walks through each layer with concrete numbers so you can validate your design before ordering hardware.

Architecture Overview

The reference node consists of four primary layers:

Camera layer: 8 PoE IP cameras connected to a managed PoE switch.
Network layer: A managed PoE switch with VLAN segmentation separating camera traffic from management traffic.
Compute layer: An edge AI compute module (Jetson Orin NX or AGX Orin) connected to the switch via Gigabit Ethernet.
Storage layer: NVMe SSD for video ring buffer, inference logs, and model storage.

The compute node pulls RTSP streams from the cameras over the local switch, performs inference, writes results and video segments to local storage, and forwards alerts or aggregated metadata upstream over a separate WAN link. No raw video leaves the site under normal operation — only inference outputs and compressed event clips.

Compute Sizing

Eight simultaneous camera streams at 1080p 15 fps requires multi-stream hardware video decode. Software decode on a CPU is not viable at this stream count — it will consume all CPU resources before inference begins.

Jetson Orin NX (16 GB) supports up to 8 simultaneous decode sessions via its dedicated NVDEC hardware. AGX Orin supports 16. For 8 cameras, Orin NX is the minimum viable compute; AGX Orin provides headroom for higher resolution inputs, more complex models, or future stream count growth.

Inference load estimate for 8-camera object detection at 1080p 15 fps using a YOLOv8s-class model in TensorRT:

Per-stream inference: ~4 ms at INT8 on Orin NX
8 streams pipelined: GPU utilization 60–80% on Orin NX
Headroom for tracking, re-ID, or classification: limited on Orin NX; comfortable on AGX Orin

If your pipeline includes secondary classification or person re-identification on detected objects, size up to AGX Orin. See the guide on RAM sizing for edge inference for memory requirements across model configurations.

Networking and PoE Layout

Bandwidth requirements for 8 cameras at 4 Mbps each: 32 Mbps aggregate. At 8 Mbps per camera: 64 Mbps. A Gigabit uplink from the PoE switch to the compute node provides ample headroom. The PoE switch itself should have a Gigabit uplink port for the compute node and 8+ PoE ports for cameras.

VLAN segmentation is strongly recommended. Place all camera traffic on a dedicated VLAN (e.g., VLAN 10) with no internet access. The compute node sits on both the camera VLAN (for stream ingestion) and the management VLAN (for remote access and alert forwarding). This prevents cameras from being reachable from outside the site and isolates camera firmware vulnerabilities from the broader network.

For detailed VLAN configuration and switch selection guidance, see networking for edge AI deployments.

PoE Power Budget

PoE cameras typically draw 5–15W per camera depending on resolution, IR illuminators, pan-tilt mechanisms, and onboard heating. A conservative estimate for outdoor fixed cameras is 10–12W each.

For 8 cameras at 12W each: 96W camera load. Add switch CPU and management overhead (~10–15W): total PoE switch power budget requirement is approximately 110–120W from the PoE power source. Select a switch with a minimum 120W PoE power budget, and prefer 150W+ for margin and future camera upgrades.

PoE standards by power class:

PoE (802.3af): up to 15.4W per port
PoE+ (802.3at): up to 30W per port
PoE++ (802.3bt): up to 60W or 90W per port

For cameras without motorized PTZ or integrated heaters, PoE (802.3af) is sufficient. For PTZ cameras or cameras with built-in IR illuminators, verify PoE+ compatibility. For full UPS and power design considerations, see power and UPS for edge deployments.

Storage Layout

Storage on the compute node serves three purposes: video ring buffer, inference output storage, and model and application storage.

Recommended partition layout for a 2 TB NVMe:

OS and applications (50 GB): Root filesystem, JetPack, Docker images, application binaries.
Model storage (20 GB): TensorRT engine files, ONNX exports, versioned model artifacts.
Inference log and metadata (30 GB): Structured event logs, bounding box records, alert history.
Video ring buffer (1.8 TB): Rolling video segments, overwriting oldest content when full.
Unpartitioned reserve (100 GB): Over-provisioning for SSD wear leveling. Leave unallocated.

At 4 Mbps × 8 cameras, the video ring buffer retains approximately 10 days of footage. At 8 Mbps × 8 cameras, approximately 5 days. Adjust bitrate and segment retention to match operational requirements. For detailed ring buffer design and write endurance math, see storage layout and ring buffer design for edge AI.

Compute Options for 8-Camera Deployments

Platform	TOPS	Max Decode Streams	RAM	TDP	Est. Cost (Module)	8-Camera Fit
Jetson Orin NX 8GB	70	6	8 GB	10–20W	~$400	Marginal (6 hw decode streams)
Jetson Orin NX 16GB	100	8	16 GB	10–25W	~$500	Suitable for 8 cameras
Jetson AGX Orin 32GB	200	16	32 GB	15–40W	~$700	Comfortable; headroom for growth
Jetson AGX Orin 64GB	275	16	64 GB	15–60W	~$900	Overkill for 8 cameras; justified for complex pipelines
RK3588 SBC	6 (NPU)	8 (HW decode)	8–16 GB	10–15W	~$200	Viable for lighter models; NPU TOPS is a bottleneck

Physical Wiring Considerations

PoE cameras require Cat5e or Cat6 cable runs. Maximum PoE cable run length is 100m (328 ft) per the 802.3 standard. Beyond 100m, power delivery degrades and the link may not negotiate PoE correctly. Use a PoE extender or a secondary PoE switch closer to distant cameras if cable runs exceed this limit.

In outdoor installations, use shielded Cat6 (STP) with weatherproof outdoor-rated jacket. Ground the shield at one end only to avoid ground loops. Conduit is recommended where cables are exposed or run across rooflines.

Label every camera port on the switch with the camera's physical location and RTSP stream IP. A small laminated label on the patch panel saves significant time during troubleshooting at 2 AM.

Common Pitfalls

Undersizing the PoE power budget: A switch with 8 PoE ports and only 65W total budget cannot power 8 cameras simultaneously at full draw. Always verify the switch's total PoE power budget, not just per-port rating.
Forgetting camera RTSP authentication: Default credentials on IP cameras are a common attack vector. Change all camera passwords before network deployment and isolate cameras on a dedicated VLAN with no internet access.
Using consumer NVMe for continuous 8-stream recording: 8 cameras at 4 Mbps each write 172 GB/day. A consumer drive will exhaust its TBW rating in 2–3 years. Use prosumer or enterprise NVMe for this write load.
Not planning for NTP synchronization: Event correlation across 8 camera streams requires synchronized timestamps. Configure NTP on the compute node and ensure cameras sync to the same NTP source.
Single point of failure on the PoE switch: If the PoE switch fails, all 8 cameras go offline. For high-availability requirements, consider a redundant switch with failover or at minimum a spare switch on-site.
Skipping thermal validation at ambient load: An 8-stream pipeline running continuously generates sustained thermal load on the compute node. Validate enclosure thermals at maximum ambient temperature before finalizing hardware.

FAQ

Can I use Wi-Fi cameras instead of PoE?

Technically yes, but not recommended for production. Wi-Fi introduces variable latency, interference, and connectivity gaps that complicate real-time inference pipelines. PoE provides deterministic bandwidth and eliminates the need for camera-side power supplies.

What RTSP stream resolution should I configure?

1080p at 15 fps is a common balance for most object detection tasks. Higher resolution (4K) increases bandwidth and decode load significantly. Configure a sub-stream at lower resolution for inference and the main stream for recording if your cameras support dual streams.

Do I need a separate NVR if I have a Jetson doing inference?

Not necessarily. A Jetson running DeepStream can handle both inference and video recording to local NVMe simultaneously. Whether you need a separate NVR depends on retention requirements, redundancy needs, and whether you need ONVIF-compatible recording for compliance.

How do I handle cameras that go offline temporarily?

Build reconnection logic into your pipeline. DeepStream supports stream reconnection natively. Design inference pipelines to continue processing active streams when one stream drops, rather than halting the entire pipeline.

What is the minimum switch port speed for camera traffic?

Gigabit (1000BASE-T) is the standard for all modern PoE cameras. 100 Mbps is technically sufficient for most bitrates but leaves no headroom and does not support the PoE+ power profile reliably on all hardware.

Should the compute node be on the same switch as the cameras?

Yes. Place the compute node on the same managed switch as the cameras but on a separate VLAN. This keeps camera traffic local (no router hop for video), while VLAN isolation maintains security boundaries between the camera network and management network.