8-Camera Edge AI Deployment Blueprint

Overview

When to use this blueprint

• Deploying 6–10 IP cameras feeding a single on-premise edge compute node
• Running continuous inference (object detection, tracking, classification) on live streams
• Environments requiring local data retention — no cloud dependency for core operation
• Locations with constrained power (sub-500W total system budget) or limited rack space
• Projects where long-term maintainability, drive endurance, and UPS coverage matter

Key constraints

• Power: PoE switch budget, compute TDP, and UPS headroom must all be planned together
• Thermals: Fanless or semi-fanless compute in a sealed enclosure requires careful thermal derating
• Bandwidth: 8 × 1080p streams at 4–6 Mbps each generates 32–48 Mbps — uplink and internal switching capacity matters
• Storage endurance: Continuous 24/7 writes to SSD from ring buffer workloads can exhaust low-endurance drives within months
• Maintenance windows: Edge deployments often have infrequent on-site access — design for remote management from day 1

Architecture Diagram

  [Camera 1–8]
       |
       | (PoE, Cat6)
       v
  [PoE Switch]  <-- managed, VLAN-capable, 150–250W PoE budget
       |
       | (Gigabit uplink)
       v
  [Edge Compute Node]  <-- Jetson-class or x86, NVMe + optional HDD/NAS
       |
       | (Ethernet)
       v
  [Router / Firewall]
       |
       v
  [Optional: Cloud / Remote Monitoring]
      

• Cameras: IP cameras with PoE power; source of all video streams
• PoE Switch: Powers cameras over Cat6 and aggregates traffic onto a Gigabit uplink
• Edge Compute Node: Runs inference, records video to local storage via ring buffer
• Router / Firewall: Network boundary for outbound traffic, VPN access, and security isolation
• Optional Cloud: Metadata sync, alert forwarding, or remote OTA — not required for core operation

Bill of Materials (BOM)

Characteristics are specified rather than exact models to remain vendor-neutral. See official vendor specs and standards when cross-checking datasheets.

Power & PoE Budget Sizing

Example calculation for 8 cameras at 12W each:

  Camera load:      8 × 12W = 96W
  Switch overhead:         10W
  Subtotal:               106W
  Headroom (25%):     +26.5W
  Recommended budget:   ≥133W  →  choose 150W or 180W tier
      

Use the PoE Power Budget Calculator to compute this for your exact camera wattage and headroom preference.

Per-port vs total budget: A switch rated "30W per port" with 8 PoE ports does not have a 240W PoE budget. The total budget is a separate, typically lower value listed in the spec sheet. Both limits must be satisfied simultaneously.

Storage Sizing (Retention + Endurance)

Retention example

Assuming 8 cameras at 1080p, H.265, average 4 Mbps per stream:

  Per camera per day:  4 Mbps × 86400 sec / 8 = ~43 GB/day
  8 cameras per day:   8 × 43 GB = ~344 GB/day
  7-day retention:     344 × 7  = ~2.4 TB
  14-day retention:    344 × 14 = ~4.8 TB
  30-day retention:    344 × 30 = ~10.3 TB
      

A 2 TB NVMe handles 7-day retention with room to spare. For 30-day retention, use an HDD archive tier and keep only 3–7 days on NVMe. A ring buffer strategy overwrites oldest footage automatically once capacity is reached.

Endurance estimate

At ~344 GB/day writes, a drive accumulates ~125 TB/year. A 600 TBW drive provides roughly 4.8 years of write life at this rate — adequate with margin. A 300 TBW consumer drive may exhaust in ~2.4 years. See SSD endurance ratings for TBW/DWPD guidance.

Networking Notes (Practical)

• Camera VLAN: Isolate all cameras on a dedicated VLAN. This prevents cameras from accessing the corporate network, reduces broadcast traffic, and simplifies firewall rules.
• Uplink capacity: 8 × 4 Mbps streams = ~32 Mbps minimum. A Gigabit uplink from switch to compute provides more than adequate headroom. Avoid 100 Mbps uplinks if you plan to record at higher bitrates or add cameras later.
• Compute-to-router link: The edge compute node should connect to the router/firewall on a management VLAN separate from the camera VLAN. This keeps inference traffic local and only forwards alerts or metadata upstream.

See Edge AI Networking: VLANs, PoE, and Bandwidth Math for a full walkthrough.

Deployment Checklist (Day 0 → Day 2)

• Day 0 — Hardware: Verify all PoE ports negotiate at expected wattage using switch management interface
• Day 0 — Hardware: Label all cables with port assignments before closing enclosure
• Day 0 — Hardware: Confirm UPS powers on cleanly and compute node performs a clean shutdown on simulated power loss
• Day 0 — Software: Update OS, JetPack/driver stack, and all inference dependencies before going live
• Day 0 — Software: Configure camera VLAN and verify cameras cannot reach the internet directly
• Day 0 — Software: Set up SSH key access; disable password auth
• Day 1 — Burn-in: Run inference pipeline at full load for 24–48 hours; monitor CPU/GPU temperature under sustained workload
• Day 1 — Burn-in: Verify ring buffer is rotating correctly and storage usage stabilizes at expected level
• Day 1 — Burn-in: Confirm all camera streams are available after a simulated power outage and restore
• Day 2 — Operations: Set up monitoring for CPU/GPU temps, disk usage, and inference latency (Prometheus, Grafana, or similar)
• Day 2 — Operations: Enable automated OS security updates or schedule regular update windows
• Day 2 — Operations: Document firmware versions, IP addresses, and hardware serial numbers
• Day 2 — Operations: Configure log rotation to prevent log writes from consuming ring buffer storage
• Day 2 — Operations: Test remote access (VPN or jump host) from an off-site location before declaring production-ready

Variants

Budget variant

• Cameras: 1080p PoE (802.3af, ~6–8W), no PTZ
• Switch: Unmanaged or basic managed PoE+, 120W total budget
• Compute: Jetson Orin Nano 4GB or entry x86 mini PC (no dedicated NPU)
• Storage: 1–2 TB consumer NVMe (verify TBW ≥ 360 TBW for 3-year margin)
• UPS: 750 VA line-interactive, basic USB management
• Tradeoff: Lower inference throughput, less visibility into PoE faults, shorter drive life at high bitrates

Balanced variant

• Cameras: 4MP PoE+ (802.3at, 10–12W), optional IR
• Switch: Managed PoE+ with VLAN support, 150–180W total budget, Gigabit uplink
• Compute: Jetson Orin NX 16GB or Orin Nano 8GB
• Storage: 2 TB NVMe (≥600 TBW) + 6 TB NAS-rated HDD for archive
• UPS: 1000 VA pure sine wave, NUT or SNMP management
• Tradeoff: Good balance of cost, inference headroom, and operational visibility

Industrial variant

• Cameras: 4MP–8MP, IP67-rated, PoE++ capable, wide operating temp range
• Switch: DIN-rail industrial managed PoE, −40°C to 75°C rated, 250W+ budget
• Compute: AGX Orin or ruggedized x86 platform with redundant storage
• Storage: Industrial NVMe with pSLC NAND (≥1200 TBW), RAID-1 or mirrored NAS
• UPS: 1500 VA pure sine wave with extended battery module, SNMP managed
• Tradeoff: Significantly higher cost; justified for 24/7 outdoor or factory deployments with multi-year SLAs

Related Guides

• 8-Camera Edge AI Reference Architecture
• PoE Switch Power Budget for 8 Cameras
• Jetson Deployment Checklist: From Unboxing to Production
• Edge AI Hardware & Infrastructure Guide