Advanced UPS Controls for AI Workloads Management: Enabling Stable Power Protection in High-Performance Computing Environments

Overview

The note starts from the load profile: representative AI cycles swing from about 10% idle to 80-100% roughly every minute, with sub-second bursts reaching 120-130% of GPU nominal power. That load pushes the battery into constant micro-cycling and stresses the upstream utility or generator. Two complementary firmware features are introduced: one that raises the DC-link voltage so short load steps are carried by the DC capacitors instead of the batteries, and input power smoothing that uses the battery as a buffer.

Input power smoothing works with a user-set fluctuation range FR%. In the example, a 1000 kW UPS at 900 kW average with FR 20% lets the input vary between 720 and 1080 kW. The example state-of-charge window is nominal 85%, minimum 50% and maximum about 97%; energy below the minimum is reserved for backup, and on generator the minimum can be lowered so smoothing may continue down to about 10% SOC at AI start-up.

Battery sizing results are decisive. In a simulation at 1200 kW, 30-100% load, 0.5 Hz and 50% duty, 7 Li-ion cabinets (64 A maximum recharge each) give only partial smoothing, while 14 cabinets give sustained full smoothing. On a gas-generator-fed GB300 installation, 6C-charge-capable modules reached 100% input smoothing with input variation limited to about 1% per second. The limits are rectifier current, generator load-step capability, battery charge/discharge current and SOC, and utility THDv tolerance; sizing should start from the measured load profile, not autonomy alone.

Key points

1. Representative AI load cycles from idle (~10%) to 80-100% roughly every minute, with sub-second peaks at 120-130% of GPU nominal power.
2. Input power smoothing uses a user-set FR%: a 1000 kW UPS at 900 kW average with FR 20% lets input vary 720-1080 kW.
3. Example SOC window: nominal 85%, minimum 50%, maximum ~97%; energy below the minimum is reserved for backup, and on generator smoothing can continue to ~10% SOC.
4. Simulation at 1200 kW, 30-100% load, 0.5 Hz, 50% duty: 7 Li-ion cabinets (64 A max recharge each) give partial smoothing, 14 give sustained full smoothing.
5. With 6C-charge-capable modules on a gas-generator-fed GB300 installation, 100% input smoothing was reached with input variation limited to about 1% per second.
6. Lab test on a large Li-ion UPS: load steps of 31-97% and 63-97% at 1 Hz kept input current nearly constant.
7. Smoothing is bounded by rectifier current, generator load-step capability, battery charge/discharge current and SOC, and utility THDv tolerance.

Using it in design

• Add an AI load-profile object to IT loads (step amplitude %, repetition Hz, duty cycle, peak % of nominal) and use it in UPS and battery sizing alongside autonomy.
• Give UPS catalogue entries supportsBatteryShield, supportsInputPowerSmoothing, minimum FR% and overload capability, and battery entries maximum charge C-rate and recharge current per cabinet.
• When the upstream source is a generator or on-site gas turbine, compare its kW/s load-step tolerance with the smoothed input profile and warn if battery charge acceptance cannot sustain smoothing.
• Reserve a backup SOC floor in runtime calculations.

Who should read it

Electrical designers sizing UPS, batteries and upstream sources for GPU clusters.

Related standards

• IEC 62040-3
• ISO 8528-5

Publication details

Authors

Vertiv

Publisher

Vertiv

Year

2025

Document number

SL-80384

Document language

English

Source type

Manufacturer

Access

Free to read

Licence

Copyright retained; citation only — the publisher’s terms apply to the original document

Citation