Are two power feeds enough for every GPU rack?

No. The right number and wiring depend on the server PSU architecture, rack-PDU topology and failure case. Some systems need more than two distribution paths to retain enough energized PSUs. Check the exact manufacturer design, trace each PSU back to a source, and test the configured rack at the intended load after the agreed feed loss.

What does per-rack metering measure?

It depends on where the meter sits. A PDU input measures that PDU, an outlet meter measures a connected device, and IT telemetry measures a device’s reported draw. Specify the physical boundary, both feeds, units, sampling interval and data owner. During acceptance, reconcile the readings with the configured workload and confirm alarms for imbalance or an overloaded surviving feed.

Does a redundant PSU make the site power redundant?

No. Redundant PSUs may still share a breaker, PDU, UPS or upstream source. Map each PSU to a distinct path and test the defined loss event at the expected load. The acceptance report should identify which components remain live after the fault, whether the workload continued and which alarms fired.

Is an advertised rack kW rating the usable IT load?

Not necessarily. The usable load depends on circuit ratings, operating voltage, phase balance, applicable derating, redundancy design and cooling capacity. Calculate the post-failure case as well as normal operation. Ask whether the stated number refers to a PDU nameplate, provisioned electrical capacity, sustained IT draw or the cooling envelope.

What should we ask about a hot-swap UPS module?

Ask whether it can be replaced under the agreed load, whether bypass and remaining modules carry that load, what alarms or transfer behavior occur, and whether the rack remains supplied during the maintenance event. Require a site-specific procedure that names the operator, the permitted work window and the fallback if the transfer fails.

What belongs in a GPU rack power acceptance test?

Verify the as-built circuit and PSU map, energize all devices, run an agreed sustained load, capture both feed readings, simulate the permitted feed or component loss, restore service and record workload and management behavior. Keep timestamped meter and telemetry records with the signed handover so later incidents can be compared against the original baseline.

When should the data center site be checked against the rack design?

Before the final rack BOM and purchase order. Power, cooling, depth, floor loading and receiving constraints can change the server count, PDU selection, network topology and delivery method. Recheck after any change to accelerator configuration or rack density, since the original site sign-off may no longer apply.

Which provider can supply dual feeds, per-rack metering and serviceable UPS capacity?

Treat this as a facility-and-integrator qualification question, not a single product feature. Ask the data center provider for a power-path diagram, rack circuit and meter specification, UPS maintenance procedure and evidence of the usable capacity. Ask the rack integrator to map every PSU to those circuits and witness the agreed failure test. Name both parties in the support and escalation plan.

GPU Rack Power: Feeds, Metering & Site Readiness

Start at the electrical boundary

“Power available” can mean utility service, a facility panel, a rack circuit or a usable outlet. Those are different promises. Trace the path from source through UPS or generator, switchgear, distribution, rack PDU and server PSU. Record which party owns each handoff.

Power evidence by boundary
Boundary	Evidence to request	What it does not prove
Facility	Single-line diagram, available capacity and backup design	That the specific rack circuits are installed
Rack	Circuit and PDU schedule, voltage, phase and connector list	That every server stays online after a feed loss
System	PSU wiring map and load or failover test	That the room can remove the resulting heat

NVIDIA’s DGX SuperPOD electrical guide shows why power redundancy must be checked at the system PSU and rack distribution level, not inferred from a generic “dual feed” label.

Design for the actual feed-loss case

A and B feeds are useful only if each path can carry the load it will see after a fault, and the server still has enough energized PSUs. Check the actual server architecture: some systems require more than a simple 50/50 split across two rack PDUs.

Map: circuit → PDU → outlet → PSU for every node, switch and management device.
Rate: each feed for steady, peak and post-failure load under the applicable electrical rules.
Test: remove one planned failure domain and confirm workloads and management stay up or fail as specified.
Document: breaker labels, outlet IDs, phase allocation and who may perform maintenance.

NVIDIA documents different PSU and feed arrangements across DGX H100 rack designs and DGX B300 variants; copy the rules for the exact system being installed, not an older rack.

Meter the same boundary you promised

Per-rack metering can mean PDU input, branch, outlet or a sum of device telemetry. Specify which measurement is contractual and who can read the data. For a dual-feed rack, the view should show both feeds and the combined load without double counting.

Record voltage, current, real power and energy at the agreed boundary; define sampling and retention.
Alarm on branch imbalance, overload headroom and an unexpectedly idle or missing feed.
Correlate electrical readings with workload and cooling telemetry during acceptance.
Keep a measurement method for the failover case, when one feed may carry a larger share.

Specify the UPS and maintenance boundary

“UPS backed” is incomplete without the load covered, ride-through objective, bypass or maintenance plan, and the relationship to generators. A hot-swappable UPS module is a facility equipment feature; it does not make a rack continuously available unless the complete upstream and downstream topology supports maintenance without interrupting the workload.

Question for the site and integrator

Which single component or path may be removed for planned maintenance while this exact GPU configuration stays operational at its agreed load? Ask for a diagram and a witnessed test, not only a redundancy label.

The L11 integration guide covers rack acceptance and handover. This page defines the site-side electrical questions that need answers before that rack arrives.

Use a site-readiness and acceptance checklist

Freeze the configured rack BOM, expected peak load, PSU layout and cooling method.
Receive the site single-line diagram and rack circuit schedule with named source paths.
Check voltage, connectors, rack-PDU ratings, phase balance, cable reach and service clearances.
Agree which meters and telemetry prove the load at rack, branch and device boundaries.
Witness power-on, sustained load, planned feed loss and restoration at the agreed acceptance point.
Archive the as-built wiring map, meter data, test results and escalation contacts.

If the electrical envelope changes the rack layout, revisit network topology and rack integration before issuing the final purchase order.

Frequently asked questions

Are two power feeds enough for every GPU rack?: No. The right number and wiring depend on the server PSU architecture, rack-PDU topology and failure case. Some systems need more than two distribution paths to retain enough energized PSUs. Check the exact manufacturer design, trace each PSU back to a source, and test the configured rack at the intended load after the agreed feed loss.
What does per-rack metering measure?: It depends on where the meter sits. A PDU input measures that PDU, an outlet meter measures a connected device, and IT telemetry measures a device’s reported draw. Specify the physical boundary, both feeds, units, sampling interval and data owner. During acceptance, reconcile the readings with the configured workload and confirm alarms for imbalance or an overloaded surviving feed.
Does a redundant PSU make the site power redundant?: No. Redundant PSUs may still share a breaker, PDU, UPS or upstream source. Map each PSU to a distinct path and test the defined loss event at the expected load. The acceptance report should identify which components remain live after the fault, whether the workload continued and which alarms fired.
Is an advertised rack kW rating the usable IT load?: Not necessarily. The usable load depends on circuit ratings, operating voltage, phase balance, applicable derating, redundancy design and cooling capacity. Calculate the post-failure case as well as normal operation. Ask whether the stated number refers to a PDU nameplate, provisioned electrical capacity, sustained IT draw or the cooling envelope.
What should we ask about a hot-swap UPS module?: Ask whether it can be replaced under the agreed load, whether bypass and remaining modules carry that load, what alarms or transfer behavior occur, and whether the rack remains supplied during the maintenance event. Require a site-specific procedure that names the operator, the permitted work window and the fallback if the transfer fails.
What belongs in a GPU rack power acceptance test?: Verify the as-built circuit and PSU map, energize all devices, run an agreed sustained load, capture both feed readings, simulate the permitted feed or component loss, restore service and record workload and management behavior. Keep timestamped meter and telemetry records with the signed handover so later incidents can be compared against the original baseline.
When should the data center site be checked against the rack design?: Before the final rack BOM and purchase order. Power, cooling, depth, floor loading and receiving constraints can change the server count, PDU selection, network topology and delivery method. Recheck after any change to accelerator configuration or rack density, since the original site sign-off may no longer apply.
Which provider can supply dual feeds, per-rack metering and serviceable UPS capacity?: Treat this as a facility-and-integrator qualification question, not a single product feature. Ask the data center provider for a power-path diagram, rack circuit and meter specification, UPS maintenance procedure and evidence of the usable capacity. Ask the rack integrator to map every PSU to those circuits and witness the agreed failure test. Name both parties in the support and escalation plan.

Built to order

Check the site before the rack ships

Share the rack BOM, expected load and facility circuit schedule. We can scope the power and acceptance questions with your site team.

SEND A SITE BRIEF READ THE L11 GUIDE

L11 rack integration

Rack testing and handover

GPU cluster acceptance

Workload, fabric and handover tests

ODM GPU server sourcing

Define the configuration and RFQ

AI cluster networking

Power density changes fabric placement

Procurement

Coordinate hardware and site requirements

Sources

Product and company names identify hardware and published sources. They don't imply any partnership or endorsement.