Power_
UPS & Battery Energy Storage for AI Halls: Ride-Through for Spiky GPU Loads
This article provides field-proven steps for sizing and integrating UPS and BESS to absorb millisecond-to-minute power transients from synchronized GPU training clusters in AI halls, including load characterization, connection sequencing, and commissioning checks performed by crews such as Leviathan Systems.
Key facts
- GPU training clusters produce simultaneous power steps when all devices enter collective operations such as all-reduce, requiring the upstream power path to absorb the transient without voltage collapse.
- UPS systems are sized for the full rack IT load plus a margin defined by the OEM specification to cover the initial period of a utility disturbance until generator or BESS takeover.
- BESS augments UPS when the required ride-through exceeds the battery autonomy of the static UPS, typically by discharging into the DC bus or AC bus via bidirectional inverters.
- NFPA 855 governs installation spacing, fire detection, and thermal-runaway mitigation for stationary battery energy storage systems above defined energy thresholds.
- All high-current DC and AC connections in the UPS-BESS path must follow torque values listed in the OEM termination kit and be verified with a calibrated torque wrench before energization.
- An infrared scan of bus joints and battery terminals is performed under load during commissioning to identify high-resistance points before they cause thermal events.
- Factory-terminated battery strings are used for BESS; field work consists of series-parallel interconnection, rack-level fusing, and insulation-resistance testing with a megohmmeter.
Characterizing synchronized GPU power transients
GPU training runs generate step changes in facility load when kernels across thousands of devices execute collective communication at the same timestep. These steps appear as simultaneous current ramps on the AC feed to each rack. The duration of the largest steps is typically under one second, yet the amplitude can exceed the steady-state draw enough to trip upstream breakers if the source impedance is too high.
Measure the actual load profile at the rack PDU during a representative training job rather than relying on nameplate values. Record three-phase current, voltage, and power factor at millisecond resolution for at least one full epoch. Compare the captured waveform against the UPS transfer time and the BESS inverter response time to determine which device must handle the leading edge of the transient.
UPS capacity and autonomy selection
Size the UPS to the sum of all connected IT loads plus the largest expected step change, using the continuous and overload ratings published by the UPS manufacturer. The battery string must support that load for the time required for the facility generator or BESS to assume the burden. In practice this means selecting a UPS whose DC bus can accept a parallel BESS connection without control-loop instability.
Verify that the UPS rectifier can recharge the combined UPS-plus-BESS batteries within the time window allowed by the site operating procedures. If the recharge interval exceeds the maintenance window, increase rectifier capacity or add a separate charger for the BESS.
BESS integration for extended ride-through
Connect the BESS inverter to the same AC bus as the UPS output or to the UPS DC bus, following the single-line diagram approved by the electrical engineer of record. The inverter must be configured with a droop or isochronous control mode that allows it to pick up load within milliseconds of a voltage dip. Set the BESS state-of-charge limits so that the system retains enough energy for the next expected utility event after a discharge cycle.
Route DC cabling from BESS racks to the inverter with the minimum number of parallel runs required by the ampacity tables in NFPA 70. Maintain manufacturer-specified bend radii and support intervals; any deviation increases inductance and slows current rise time during a step load.
High-current connection and torque verification
Terminate all AC and DC conductors using the lugs and hardware supplied in the OEM termination kit. Apply the torque value listed on the equipment nameplate or installation manual; under-torque creates high-resistance joints that heat under load, while over-torque can crack terminal pads. Re-torque after the first thermal cycle if the manufacturer requires it.
Perform a low-resistance measurement across every joint with a micro-ohmmeter before energization. Record values and compare against the acceptance criteria in the commissioning plan; any joint exceeding the threshold must be disassembled, cleaned, and re-terminated.
Common failure modes during commissioning
The most frequent field failure is a control mismatch between the UPS and BESS inverters that causes oscillation or delayed pickup during a step load. This occurs when firmware versions or droop settings are not aligned; catch it by performing a controlled load-step test with a resistive load bank before connecting live GPU racks.
Another recurring issue is insulation breakdown on DC battery strings after installation. Moisture ingress or mechanical damage during rack movement creates ground faults that only appear under voltage. Always perform a megger test on each string after mechanical installation and again after the first charge cycle. A third failure is thermal runaway propagation when BESS modules are installed with insufficient aisle spacing or blocked exhaust paths; confirm clearances against the NFPA 855 layout drawing and verify airflow sensors are functional before placing the system online.
Commissioning sequence and acceptance testing
Energize the UPS first on utility, confirm output voltage and frequency stability, then bring the BESS online and verify synchronization. Apply a step load equal to the largest expected GPU transient using a load bank; record UPS and BESS response times, voltage deviation, and recovery. Only after the power path passes this test should GPU racks be connected.
Document all setpoints, firmware revisions, and test results in the commissioning package.
Standards referenced: NFPA 855 · NFPA 70 · IEEE 1547
Frequently asked_
How much BESS energy is required when UPS runtime is limited?
Calculate the energy by multiplying the measured step-load magnitude by the additional seconds of support needed beyond the UPS autonomy, then apply round-trip efficiency and depth-of-discharge limits from the BESS manufacturer. Add a margin for cell aging. Verify the resulting kWh value against the inverter continuous-discharge rating so the BESS can actually deliver the power, not just the energy.
What torque value should be used on BESS DC bus bars?
Use the exact figure stamped on the terminal or listed in the OEM installation manual for that specific lug and bus-bar material combination. After the first thermal cycle, re-check torque if the manual requires it. Never substitute a generic torque chart.
Can the same BESS support both UPS ride-through and generator-start bridging?
Yes, provided the control system is programmed with two distinct discharge curves and the state-of-charge reserve for generator cranking is protected. The inverter must be sized for the combined peak of the GPU step plus any motor loads on the generator bus.
Why does infrared scanning occur under load rather than at rest?
High-resistance joints only produce measurable temperature rise when current flows. A scan performed with the system energized at design load reveals problems that would remain invisible during a de-energized inspection.
Who performs the final integration test between UPS, BESS, and GPU racks?
The electrical contractor or specialized commissioning crew such as Leviathan Systems executes the integrated test after all subsystems have passed individual acceptance. The test must be witnessed by the owner or their representative and the results added to the as-built documentation.