Power_
Harmonics & Power Quality at GPU Density: Taming the Switch-Mode Load
Details the mechanisms by which dense GPU racks produce harmonic currents through switch-mode supplies, the resulting waveform distortion on facility feeders, and the ordered steps for measurement, filtering, and inrush control that field crews apply during rack integration.
Key facts
- GPU power supplies are switch-mode units whose front-end rectifiers produce odd-order harmonics, primarily 3rd, 5th, and 7th.
- IEEE 519 defines recommended limits for voltage and current distortion at the point of common coupling.
- High total harmonic distortion of current on a shared feeder can cause neutral conductor overheating and transformer derating.
- Inrush current in GPU racks occurs at PSU startup and is limited by soft-start circuits or sequenced breaker closure per OEM rack documentation.
- Active harmonic filters inject counter-phase current and are sized to the measured THDi rather than to rack kW rating.
- Crews perform pre-energization harmonic surveys with calibrated three-phase power quality analyzers before final PDU terminations.
- MPO and fiber domains carry scale-out fabrics only; power quality work remains entirely on the AC distribution side and does not interact with NVLink copper backplanes.
Switch-mode supply behavior under GPU load
Each GPU tray contains multiple AC-DC supplies whose diode or active front-end rectifiers draw current only near the peaks of the voltage sine wave. The resulting pulsed current contains strong odd harmonics that add arithmetically on the neutral and circulate in delta-wound transformers.
Because rack densities are high, the aggregate harmonic current can exceed the limits that facility transformers and conductors were originally sized to carry. The distortion also raises RMS current without a corresponding increase in real power, lowering true power factor.
Field crews therefore treat the rack power shelf as a known harmonic source rather than a resistive load when planning feeder and transformer capacity.
Waveform distortion on shared distribution paths
When multiple racks share a PDU or busway, the summed harmonic currents produce flat-topped voltage waveforms at the rack inlet. This voltage distortion feeds back into every supply, increasing internal DC ripple and component stress.
Neutral currents rise significantly when triplen harmonics are present, requiring upsized neutrals or separate harmonic mitigation on each three-phase leg.
Crews record phase-to-neutral voltage THDv at the rack whip before and after load ramp to confirm the facility bus remains within acceptable limits.
IEEE 519 application at the rack level
IEEE 519 sets both voltage distortion limits at the point of common coupling and current distortion limits based on short-circuit ratio. In practice the rack is treated as an individual load whose contribution must be measured at the PDU output to verify compliance before the facility accepts the rack.
The standard requires that individual harmonic currents stay below stated percentages of the maximum demand load current; crews therefore capture demand intervals rather than instantaneous snapshots.
When measured values exceed the calculated limit, the sequence is to add passive or active filters before proceeding to full rack population.
Harmonic mitigation hardware choices
Passive shunt filters tuned to the 5th and 7th harmonics are installed at the PDU when space and heat load allow. They reduce THDi but introduce leading power factor at light load, which must be checked against generator capability.
Active harmonic filters are preferred for variable GPU workloads because they respond dynamically and do not require retuning when rack population changes. Sizing begins with a power quality study that captures the worst-case THDi under the expected job mix.
Filter placement is always upstream of the rack branch breakers so that the filter can treat the entire rack as a single harmonic source.
Inrush sequencing during rack energization
GPU PSUs present a high inrush when their bulk capacitors charge. Sequential closure of rack-level breakers, staggered by intervals specified in the OEM documentation, keeps the combined inrush below the instantaneous trip setting of the upstream breaker.
Where the PDU supports it, crews enable the rack management controller's staged power-up feature before applying AC. This firmware sequence is verified against the PDU's own inrush specification.
If inrush still trips breakers, an external pre-charge circuit or soft-start module is added at the rack inlet; this step is performed only after confirming the PSU vendor permits external pre-charge.
Common field failure modes and detection
The most frequent issue is under-sized neutral conductors that overheat after racks reach steady state; infrared scans of terminations during the first hours of full load catch this before insulation damage occurs.
Another recurring failure is active filter resonance with existing power-factor-correction capacitors; crews verify that any existing PFC stage is disabled or detuned before filter commissioning.
Missed inrush sequencing shows up as nuisance trips during maintenance windows; the root cause is usually an operator bypassing the controller and closing all breakers manually. Pre-job checklists now require documented verification of the staged-power-up script before any rack is released to production.
Standards referenced: IEEE 519
Frequently asked_
Do we need to derate transformers when adding GPU racks?
Yes when measured THDi is high. The K-factor of the transformer must be recalculated from the actual harmonic spectrum captured during the site survey. Leviathan Systems includes this calculation in the pre-deployment power study so the facility can order K-rated or derated units before rack delivery.
Where do we place active harmonic filters relative to the rack PDUs?
Filters sit on the line side of the rack branch breakers so they see the aggregate load. Placing them downstream of individual rack PDUs leaves the PDU and its internal wiring unprotected and can create control-loop interaction between multiple filters.
How long should the power quality recording run before sign-off?
A capture period that includes both minimum and maximum expected GPU utilization is required. Shorter windows miss the load-dependent rise in 5th and 7th harmonics that appears only under sustained compute jobs.
What happens if measured THDv at the rack inlet exceeds the project target?
The rack is not accepted until the upstream voltage distortion is corrected. Common causes are existing non-linear loads on the same bus or undersized feeders; both must be addressed before the GPU racks are energized.