// Power Protection

Power Protection Architecture for Edge AI: UPS Design for Cameras, Compute, and Networking

Last updated: March 2026

Designing power protection for edge AI means protecting the entire pipeline — compute, PoE switch, cameras, and networking — on the same UPS circuit. This page covers system architecture, not detailed battery math. For UPS capacity formulas and runtime calculation, see the dedicated UPS sizing guide.

Protect compute + switch + cameras

Graceful shutdown automation

5–10 min shutdown target

UPS must protect full pipeline

Quick Answer

Power protection is a system architecture problem. You must protect the full pipeline — compute, PoE switch, cameras, and networking infrastructure — on the same UPS circuit. If only the compute stays up while the switch loses power, the cameras disconnect and inference fails anyway.

This page focuses on designing resilient UPS architecture: what to protect, how to automate graceful shutdown, testing strategies, and deployment patterns. For detailed UPS capacity calculation, battery sizing formulas, and runtime math, see the UPS Sizing for Edge AI Systems page.

Key steps: (1) measure total sustained load across all components, (2) design graceful shutdown with NUT automation, (3) test shutdown behavior under battery, (4) plan for maintenance and battery aging.

Planning Takeaway

Architecture matters more than battery label marketing. The most common UPS failures come from protecting only the compute node while the switch or cameras remain unprotected, or deploying without NUT automation. Protect the full pipeline and test shutdown automation before production.

Who This Page Is For

Engineers deploying edge AI nodes with PoE cameras and needing power protection
Teams evaluating UPS form factors (tower, rack-mount, DIN-rail) for their deployment
Operators planning runtime targets and graceful shutdown automation
Architects designing multi-node or outdoor installations with power redundancy requirements

Why UPS at the Edge

Data center servers operate on redundant power feeds with automatic transfer switches and generator backup. Edge nodes in a retail location, a warehouse, or an outdoor cabinet are typically connected to a single utility circuit with no redundancy. Power events — brief sags, brownouts, and hard cuts — are far more common at the building level than at a colocation facility.

The consequences of unprotected power loss at an edge AI node include:

NVMe SSD filesystem corruption or write-in-progress data loss (particularly severe without power loss protection in the drive)
Docker container state corruption requiring manual recovery
Inference pipeline crash requiring on-site restart if watchdog recovery fails
Hardware damage from voltage sag on sensitive compute platforms
Loss of video footage covering the event window

A properly sized UPS eliminates all of these failure modes during outages short enough for graceful shutdown, and suppresses voltage anomalies that can damage hardware even without a full power cut.

Total Load Calculation

Measure or estimate the sustained power draw of every device on the UPS circuit:

Component	Typical Sustained Draw
Jetson Orin Nano (7W mode, active inference)	7–10W
Jetson Orin NX (15W mode)	15–20W
Jetson AGX Orin (40W mode)	40–55W
PoE switch (8-port, 120W PoE budget)	25–40W (switch itself, excluding PoE output)
PoE cameras (per camera, 10–12W)	80–96W for 8 cameras
NVMe SSD (active write)	2–5W
Ethernet media converter or router	5–15W

A typical 8-camera Jetson Orin NX node total: ~170–200W sustained. Always use measured values from a calibrated power meter rather than rated TDP when available — actual draw is often 20–30% below rated maximum.

Use the Power Budget Planner to calculate and document your total load across all components.

For component-level power breakdown, see the 8-camera reference architecture and networking for edge AI for PoE switch power specifics.

Protecting the Full Pipeline

A common mistake is protecting only the compute node while leaving the PoE switch or cameras unprotected. This creates a false sense of security: the compute node may stay up on battery, but if the network infrastructure loses power, the pipeline fails anyway.

The rule: every component that must remain operational during the graceful shutdown window must be on the same UPS circuit. This includes:

Compute node (Jetson, GPU host, ARM board) — required for inference and shutdown orchestration
PoE switch — if this loses power, cameras disconnect and the pipeline crashes regardless of compute status
Cameras — powered via PoE through the switch; if the switch is on battery, cameras remain operational
Networking infrastructure (modem, router, media converter) — if cloud-dependent systems need to notify remote services before shutdown
Storage media (NVMe, external USB) — ideally on the UPS for clean filesystem shutdown

For detailed UPS capacity calculation and runtime math, see the UPS Sizing guide. This section focuses on ensuring your architecture covers all critical components.

Failure Scenarios UPS Addresses

Hard power cut: UPS transfers to battery instantly (online UPS) or within 4–8ms (line-interactive). Either is fast enough to prevent compute platform reset.
Voltage sag (brownout): Line-interactive and online UPS regulate output voltage during sags, protecting equipment from under-voltage damage.
Voltage surge: UPS AVR (automatic voltage regulation) clamps surges before they reach connected equipment. Surge-only protection strips do not provide runtime on outage.
Brief interruption (sub-second): Most utility flicker events last under 200ms. A UPS on battery through a brief interruption prevents any equipment disruption.
Extended outage: UPS provides runtime for graceful shutdown sequence. After shutdown, the site has no power regardless — the UPS goal is clean shutdown, not indefinite operation.

UPS Form Factors for Edge

Compact tower (desktop form factor): 300–1500 VA. Fits on a shelf or in a small cabinet alongside networking gear. Most common choice for single-node edge deployments. Brands offer USB or serial communication ports for graceful shutdown automation.

1U or 2U rack-mount: For deployments in equipment racks (wiring closets, server rooms at the edge site). Higher VA ratings available. Better for multi-node installations sharing a single UPS.

Compact DIN-rail DC UPS: For industrial cabinet deployments. Takes 24V or 48V DC input and output, with an internal battery pack. Eliminates AC-DC conversion loss when the edge compute runs on DC. Relevant when the full node is already DC-powered.

Mini tower (sub-200 VA): Adequate for single Jetson Nano or Pi-class nodes without cameras. Not sufficient for nodes with PoE switches and multiple cameras.

UPS Type Comparison

UPS Type	Transfer Time	Voltage Regulation	Efficiency	Cost	Best For
Standby (offline)	4–12ms	None (pass-through)	95–98%	Low	Basic outage protection; not ideal for sensitive loads
Line-interactive	2–6ms	AVR (boost/buck)	92–96%	Medium	Most edge AI deployments; good balance of cost and protection
Online double-conversion	0ms (always on battery)	Full isolation	85–92%	High	Sensitive loads; high-uptime requirements; higher ambient noise
DC UPS (DIN-rail)	~0ms	Depends on design	90–95%	Medium–High	Industrial DC-powered edge cabinets

Graceful Shutdown Design

A UPS without graceful shutdown automation is only half a solution. When the UPS transfers to battery, the compute node must be notified so it can execute a clean shutdown before the battery is exhausted.

Most UPS units include a USB communication port. The open-source NUT (Network UPS Tools) daemon reads UPS status via USB and can trigger system shutdown when battery reaches a configured threshold (e.g., 30% remaining or 3 minutes estimated runtime left).

Shutdown sequence for an edge AI node:

UPS transfers to battery → NUT detects event → NUT notifies system
Inference pipeline receives SIGTERM → flushes in-progress video segment → saves checkpoint state
Docker containers stop gracefully
NVMe sync and unmount
System shutdown complete within 60–90 seconds

Design and test this sequence before deployment. A graceful shutdown that takes 3 minutes but only has 2 minutes of battery is not a working solution. Always test NUT-triggered shutdown under actual battery conditions before going to production.

Common Pitfalls

Sizing UPS by VA rating alone: VA is apparent power. Watts is real power. A 600 VA UPS may only support 360W real load. Verify both the VA and watt ratings against your measured node load.
Not automating graceful shutdown: A UPS without NUT or equivalent automation just delays the crash until the battery dies. Automate clean shutdown as part of initial deployment, not as a future improvement.
Protecting only the compute node: If the PoE switch or router loses power while the compute node is on battery, cameras disconnect and inference fails. Protect all pipeline components on the same UPS.
Forgetting camera PoE power in load calculation: The PoE switch draws its rated power budget from the AC outlet, not just the switch's own consumption. If cameras draw 96W total through PoE, that 96W comes from the UPS via the switch.
Not testing and maintaining deployed UPS systems: Lead-acid batteries degrade significantly after 2–3 years of 24/7 float charging. A UPS reporting "battery OK" may provide 50% of rated runtime if the battery is old. Test deployed UPS systems periodically by discharging under load and measuring actual runtime vs. calculated estimates. Replace batteries on schedule and re-validate NUT shutdown automation after any maintenance.
Undersizing for cold weather: Lead-acid and some lithium battery chemistries lose significant capacity at low temperatures. An outdoor cabinet UPS in a cold climate may deliver 70–80% of rated capacity in winter. Size accordingly.
Using consumer surge strips instead of UPS: Surge strips protect against voltage spikes but provide no runtime on power loss. They do not prevent filesystem corruption or pipeline crashes from hard power cuts.

How to Size This in a Real Deployment

Measure real sustained load: Use a calibrated power meter at each component. Record the average draw during active inference over at least 10 minutes. Note peak draw (startup, thermal spike) separately.
Add PoE camera and switch overhead: Total PoE current (cameras + switch) flows through your UPS circuit. Sum the individual camera draws (typically 10–12W each) plus switch consumption (~25–40W for an 8-port managed switch).
Set runtime target: Most deployments target 5–10 minutes of runtime for graceful shutdown. If your site has automatic generator startup within 30 seconds, 2–3 minutes is sufficient. Document the target clearly.
Size the UPS for your total load and runtime: For detailed capacity formulas and runtime math, use the UPS Sizing guide. That page covers Wh calculation, VA vs watts, derating for temperature, and aging effects. For this step, specify your total load (Watts), runtime target (minutes), and climate zone (which affects battery derating).
Test NUT-triggered shutdown under battery: Before going to production, physically unplug the UPS from AC power and verify that NUT detects the event and triggers system shutdown with at least 2 minutes of runtime remaining. Document the actual runtime vs. calculated runtime and adjust your runtime target if needed.

Frequently Asked Questions

What is NUT and how does it work with a UPS?

NUT (Network UPS Tools) is an open-source daemon that communicates with UPS hardware via USB, serial, or network. It monitors battery status and can trigger system shutdown scripts when configurable thresholds are reached. Available in standard Ubuntu and Debian package repositories.

Can I use a lithium battery UPS for outdoor installations?

Lithium-based UPS units offer higher energy density, faster recharge, and better cycle life than lead-acid. They are well-suited to outdoor installations within their temperature range. Verify the battery's operating temperature range against your deployment environment's extremes.

How long should my UPS runtime target be?

5–10 minutes is sufficient for graceful shutdown in most deployments. If the site has a generator that starts automatically within 30–60 seconds, 2–3 minutes of runtime is adequate. Extended runtime (hours) requires external battery packs and significantly larger UPS units.

Should the UPS protect just the compute node or the whole installation?

Protect everything connected to the pipeline — compute node, PoE switch, and cameras — on the same UPS circuit. If the PoE switch loses power, cameras stop streaming and the inference pipeline crashes anyway, regardless of compute node power status.

What happens to in-flight video when power cuts?

Without a UPS, the current video segment file is likely corrupted. With a UPS and graceful shutdown, the pipeline can flush the in-progress segment to disk cleanly before shutdown. Some systems use memory-mapped ring buffers that survive brief power events if the UPS maintains power long enough.

Do I need a UPS for a node on a PoE injector?

Yes. PoE injectors are passive — they pass AC power to the switch and have no battery backup. A UPS upstream of the PoE injector protects the entire chain. Size the UPS to include the PoE injector's full output load, not just its idle consumption.

The Bottom Line

Power protection for edge AI is primarily an architecture and automation problem, not just a battery capacity problem. Protect the entire pipeline — compute, switch, cameras, and networking — on the same UPS circuit. Design and test graceful shutdown automation with NUT before production. Monitor battery health quarterly and plan for replacement every 2–3 years.

For UPS capacity calculation, detailed runtime formulas, Wh sizing, and derating due to temperature or aging, see the UPS Sizing guide. For broader edge AI system architecture decisions, see the 8-Camera reference architecture and networking guide.