Power_
Rack PDU Selection & Metering for GPU Racks: Outlet Count, Phase, Telemetry
Details the concrete parameters for selecting metered and switched PDUs in 50-150 kW GPU racks, including outlet counts, phase balancing, and per-outlet telemetry requirements for commissioning and ongoing operations.
Key facts
- Three-phase 208 V PDUs deliver more power than single-phase circuits at the same amperage due to the phase relationship.
- IEC 60320 C13 and C19 connectors carry continuous loads at their rated voltage per the component standard.
- Per-outlet metering in switched PDUs reports true RMS current and power for individual receptacles.
- Rack PDUs for AI workloads commonly use IEC 60309 3P+N+E plugs in 16 A, 32 A, or 63 A ratings.
- UL 489 listed breakers inside PDUs provide branch-circuit protection coordinated with upstream distribution.
- Factory-terminated power whips with pre-crimped lugs eliminate field torque errors on GPU rack terminations.
Determining Outlet Count from GPU Power Budget
Start with the nameplate power of each GPU tray plus host CPUs, memory, and networking cards, then add headroom for PSU efficiency and transient spikes per the OEM guidance. Divide total rack draw by the continuous rating of the chosen outlet type to arrive at the minimum number of receptacles per phase. Record the resulting outlet map on the rack elevation drawing before ordering PDUs so that every power supply has a dedicated circuit.
This calculation also sets the PDU frame size. Undersizing outlet count forces daisy-chaining or Y-cords that violate single-cord policies and complicate per-outlet metering.
Phase Selection and Load Balancing Rules
Three-phase PDUs are required above roughly 20 kW per rack because single-phase circuits cannot deliver the needed amperage within standard breaker sizes. Distribute GPU power supplies evenly across the three phases so that measured current on each leg stays within tolerance of the others under full load. Imbalance outside the allowed range trips upstream breakers or reduces the effective capacity of the PDU.
Measure phase currents at the PDU controller after the first full burn-in run. If one leg exceeds the others by more than the allowed tolerance, swap cords at the outlet level rather than re-terminating upstream. Document the final phase map in the as-built package for future tray swaps.
Why Per-Outlet Telemetry Is Required for GPU Racks
GPU power draw varies sharply with workload; aggregate rack metering hides individual tray faults or firmware-induced spikes. Per-outlet data lets operators correlate a specific GPU or NVLink tray with its power signature during commissioning and later troubleshooting. Telemetry also supports capacity planning by showing actual utilization versus nameplate ratings over weeks of production traffic.
Select PDUs whose metering accuracy meets the requirement in the relevant standard and that expose SNMP or Redfish endpoints for integration with DCIM platforms. Verify that the controller samples frequently enough so that short-duration spikes above the breaker trip curve are captured.
Metered-Only versus Switched PDUs for Commissioning
Switched PDUs add remote outlet control that speeds sequential power-up sequences and allows isolation of a single tray without rack shutdown. Metered-only units are smaller and cheaper when remote switching is not needed, but they still require local outlet-level current monitoring for the same telemetry value. Choose switched models when the rack contains more than 20 field-replaceable power supplies.
Both types must retain outlet state across power cycles so that a brief upstream outage does not require manual intervention. Confirm this behavior in the OEM firmware release notes before deployment.
Common Failure Modes and Field Detection
Loose IEC 60320 connectors overheat when draw exceeds a high percentage of rating for extended periods; infrared scans of the rack during burn-in catch temperature rise before insulation damage occurs. Phase imbalance from mis-mapped cords is the second most frequent issue and is detected by comparing the three leg currents reported by the PDU controller.
Factory whip terminations that were not torqued to the OEM specification loosen under vibration; always re-check torque values on the first preventive-maintenance visit. Metering drift outside acceptable limits usually indicates a failing current transformer and is caught by comparing PDU readings against a calibrated clamp meter during periodic validation.
Commissioning Sequence
After mechanical installation, connect each whip to its designated outlet, close the PDU breakers one phase at a time, and confirm zero current on unused outlets. Run the full GPU workload while logging per-outlet power; any outlet exceeding its continuous rating triggers a cord or tray review.
Export the telemetry dataset to the customer DCIM before sign-off. This baseline becomes the reference for all future change-control tickets involving power supplies or firmware updates. Leviathan Systems teams follow this sequence on every high-density deployment.
Standards referenced: IEC 60320 · IEC 60309 · UL 489 · TIA-942-B
Frequently asked_
How many outlets do I need for a 120 kW GPU rack?
Calculate total draw including headroom, then divide by the continuous rating of the chosen connector. A typical high-density rack at 208 V three-phase needs multiple PDUs fitted with sufficient C19 outlets when each GPU tray draw is known from nameplates. Confirm the exact count against the tray power supply nameplates before ordering.
Does per-outlet metering replace rack-level metering?
No. Per-outlet data is required for tray-level diagnostics and capacity planning, while rack-level metering remains the billing and alarm source for the facility. Both must be collected and time-synchronized in the DCIM platform.
What accuracy is required on PDU current sensors?
The relevant standard specifies accuracy for higher-tier sites. Verify the OEM specification sheet lists true-RMS measurement and perform a spot-check with a calibrated clamp meter during commissioning.
When should I choose switched PDUs over metered-only units?
Select switched PDUs when the rack contains more than 20 field-replaceable power supplies or when remote power sequencing is part of the standard operating procedure. Metered-only units suffice for simpler racks where physical access is always available.
How do I verify phase balance after installation?
Read the three leg currents from the PDU controller under full GPU load. Swap cords at the outlet level until the difference between any two legs meets the allowed tolerance. Record the final map in the as-built documentation.