Redundant Power Supply: A/B Rack PDU Feed Mapping and Breaker Sizing

A redundant power supply arrangement keeps a server operating after the specified power failure, provided the surviving PSU modules have enough capacity. For A/B rack PDU feed mapping and breaker sizing, first confirm that the server can survive an entire feed loss, then check the current carried by every surviving outlet bank, line conductor and upstream circuit.

The practical output is a traceable connection map and two separate checks: A lost, B carrying the load; B lost, A carrying the load. This is the review to complete before fixing PDU and server power cord quantities for a rack build or expansion.

Confirm what the redundant power supply must survive

Define the failed element before assigning cords. Losing one PSU, one outlet-bank breaker, one rack PDU or one UPS can remove different numbers of power supplies from the same server.

In N+1 module redundancy, N is the number of modules needed to support the configuration, with one additional module available. A 2N, or N+N, arrangement provides two complete capacity groups. The required count and permitted slot grouping come from the server platform and its configured operating mode.

The ASUS RS720A-E13 user manual, section 2.11, requires system consumption to remain below one PSU module's maximum output for its redundant hot-swap mode. It also notes that output varies with input voltage. Check supported DC output capacity at the actual supply conditions; use AC input demand for the facility circuit calculation.

Slot numbering also matters. The NEC Express5800/A1040d family user guide, printed page 187, groups PSU1/PSU2 into channel A and PSU3/PSU4 into channel B. Follow the documented groups for the installed platform rather than assigning alternating slot numbers by habit.

A multi-PSU exception: DGX H100

NVIDIA's DGX SuperPOD: Data Center Design Featuring NVIDIA DGX H100 Systems, Electrical Specifications, requires four of six PSUs energized. Splitting those six modules three-and-three across two feeds leaves only three energized after either feed fails. That arrangement fails the platform's operating requirement.

The guide describes three rack power paths with two PSUs per path. Losing one local path then leaves four energized. Trace those paths upstream as well: in its N+1 arrangement, two rack PDUs share one UPS, so loss of that shared UPS has a larger failure scope than loss of a single rack PDU.

Selection rule: if the proposed outage leaves fewer than the OEM-required modules or insufficient supported output, revise the feed architecture before selecting higher-current PDUs.

Map every server power cord to its upstream source

Use one record per PSU cord, with identifiers that can be checked at both ends. The connection chain is: PSU slot → cord → PDU outlet → protected bank → PDU input → panel circuit → upstream source.

Label connections with the PDU identity and breaker number; color helps technicians follow the route. NVIDIA's power-connection guidance calls for both source/breaker labels and color coding. A blue cord and an orange cord still need the written map to establish where they terminate.

Phase identity and failure-domain identity answer different questions. Two outlets protected by the same breaker share that breaker failure. Identical phase labels on independent A/B sources can belong to separate circuits. Conversely, different phases supplied by one UPS still share that UPS. Trace the actual protective devices and upstream junctions.

For each server power cord, match the approved equipment connection, PDU outlet, assembly current/voltage rating, destination-market requirements and routed length. Some platforms prescribe the supplied cordset. Use the C13 vs C19 power cord guide for the connector-side decision once the map establishes the required endpoints.

Size breakers for the surviving feed

Use the post-failure continuous load and the installed protection's applicable rating or setting. For the standard North American application described in Eaton's January 2026 application paper AP01200008E:

All currents are in amperes for the circuit being assessed. Eaton describes a continuous load as one whose maximum current is expected to persist for three hours or more. Apply the locally adopted requirements and equipment listing when selecting the installation's protection.

For an assembly listed for 100% continuous operation, the same paper gives the sum of continuous and noncontinuous currents without the 1.25 multiplier. That application depends on the complete assembly and the manufacturer's installation conditions, including its enclosure requirements. Conductors, terminations, cordsets, outlets and the PDU retain their own limits.

Use the setting, not just the frame size

Schneider Electric FAQ FA104355 gives a concrete distinction: a 150 A H-frame set to 100 A has a standard continuous-load allowance of 0.80 × 100 A = 80 A. The 150 A frame size is not the calculation base in that setting example.

Record the rating basis beside every limit. If a PDU specification already states its permissible continuous input current, use that figure directly; multiplying it by 80% again would apply a second reduction.

What the balanced A/B planning rule means

For equal-voltage feeds with approximately equal sharing and unchanged load behavior, a useful initial screen is:

Where an 80% continuous limit applies to the breaker rating, that screen becomes 40% of the breaker rating per feed in normal operation. This is a conditional planning relationship: half of the usable surviving capacity. At equality, the screen leaves no additional growth reserve.

Use explicit transfer-state demand when feeds share unequally, voltages differ, PSU standby modes change the distribution, or loads concentrate on particular banks. Final breaker selection also needs conductor and termination ratings, installation conditions, available fault current and protection coordination.

Check rack power density in both failure states

Rack power density is constrained by the most heavily loaded surviving element. Adding the A and B nameplate capacities does not establish the rack's redundant operating capacity.

For each bank, line conductor, PDU inlet and upstream circuit, record the sustained current in normal operation and after each defined failure. Evaluate the other equipment sharing those circuits as part of the same load record.

For each element and scenario, calculate:

A negative margin identifies a capacity failure; zero leaves no continuous-current reserve. Set any additional growth allowance through the project requirements. Correct the limiting bank or path by redistributing compatible connections, reducing load or revising the distribution design, then recalculate both failure states.

A rack-total reading can hide a bank overload. Derive each bank's demand from the outlets actually connected to it. “Double every bank” only works when the transfer distribution justifies that assumption. For three-phase PDUs, check each line current; unused capacity on other phases cannot clear an overloaded conductor. Line-line outlet loads also require the correct topology when translating outlet demand into line current.

Keep the transient check separate. The 80% application allowance is not an instantaneous trip threshold: Schneider's FAQ distinguishes continuous-load application from breaker trip characteristics. Review transfer amplitude and duration, restoration inrush and recovery against the actual time-current curves and platform limits. A positive sustained margin alone does not establish successful ride-through.

For broader monitoring and outlet-selection choices, continue with how to choose a rack PDU for GPU servers.

Validate the map and carry it into the BoM

Acceptance needs a physical trace and an approved failure-state verification. Agree the maintenance window, intended failure boundary, workload, monitoring points and recovery procedure with qualified facilities staff and the server owner before any interruption.

1. Match the installed hardware to the map. Confirm PSU slots, approved cords, PDU outlets, bank breakers and upstream circuit labels. Check the actual redundancy mode.
2. Establish the loaded baseline. Record the intended workload, supply conditions and bank/line currents. Retain the platform data supporting the expected maximum demand.
3. Verify each approved failure scenario. Qualified staff should carry out the platform-specific test plan while the server owner verifies continuity and records PSU status, sustained current and transfer behavior.
4. Check restoration. Review reconnection/inrush behavior, alarms and the return to the intended sharing mode. Retain the event traces and breaker settings with the capacity record.
5. Freeze the checked map for procurement. Carry exact PDU topology, outlet counts, cord endpoints, rated assemblies, routed lengths and labels into the BoM. Repeat affected checks when hardware or connections change.

The resulting map gives the AI Rack Power Kit requirements a clear basis: the correct accessory lines, allocated to identified outlets and feeds. Keep the map revision with the rack's installation and commissioning records.

Technical references

• ASUS: RS720A-E13 User Manual, V2, section 2.11, “Redundant power supply units” — single-module output limit for redundant mode and input-voltage dependence.
• NEC Corporation: Express5800/A1040d, A2040d, A2020d, A2010d User's Guide, printed page 187 — PSU channel groups and redundancy modes.
• NVIDIA: DGX SuperPOD: Data Center Design Featuring NVIDIA DGX H100 Systems, Electrical Specifications — energized PSU requirements, source topology, labeling and phase balancing.
• Eaton: UL listed 100%-rated molded case circuit breakers, Application Paper AP01200008E, January 2026, pages 1–2 — continuous-load sizing and listed-assembly conditions.
• Schneider Electric: FAQ FA104355, “What is the difference between a 80% rated breaker and a 100% rated breaker?”, modified October 14, 2024 — setting-based continuous allowance and trip-curve distinction.
• Raritan: PDU Circuit Breakers and Branch Circuits: Why Monitor Both? — outlet-bank monitoring and the limits of input-only readings.