AI Load Dynamics - A Power Electronics Perspective

Overview

The work is a 2025 arXiv preprint by Yuzhuo Li and Yunwei Li of the University of Alberta (arXiv:2502.01647). It examines how AI accelerator load transients interact with the multi-stage power conversion chain in data centers, from MV transformers and UPS to rack PSUs and voltage regulators. Modern AI accelerators can exhibit power variations exceeding 50% of TDP. The study surveys AI rack configurations: a traditional A100 rack at 32 kW (400 W per GPU), a traditional H100 rack at 34 kW (700 W per GPU), a liquid-cooled GB200 NVL36 at 73 kW and an NVL72 at 132 kW (1,200 W per GPU).

A 7392 kW IT-rack configuration corresponds to approximately 10.5 MW in total including cooling infrastructure. It also covers AC and DC power-chain architectures and their efficiencies, and PSU and VRM topologies and control. As efficiency figures, MVAC to 480 VAC transformation typically achieves about 98.5%-99.2%; extra North American stages (480 V to 277 V) add roughly 1-2% penalty per stage. Double-conversion UPSs typically attain about 94-96% at rated load and line-interactive 97-98%, but data centers may operate at only 30-50% of rated capacity, so in-service efficiency deviates from these figures.

A measurement case study shows that after a roughly 180 W GPU load step the PSU AC input current lagged by about 1-2 AC cycles (20-40 ms), up to about 50 ms. The rack PSU bandwidth of 30-50 ms bounds the fastest net swing seen upstream, meaning the final conversion stage, not the VRM, sets the fastest swing. Synchronised multi-rack swings still reach the utility interface, so additional buffers (supercapacitors, high-slew UPS, staggered scheduling) are recommended.

Key points

1. Rack configurations: traditional A100 rack 32 kW (400 W/GPU), traditional H100 rack 34 kW (700 W/GPU), GB200 NVL36 liquid-cooled 73 kW and NVL72 liquid-cooled 132 kW (1,200 W/GPU).
2. A 7392 kW IT-rack configuration corresponds to approximately 10.5 MW total including cooling infrastructure.
3. Modern AI accelerators can exhibit power variations exceeding 50% of TDP.
4. MVAC to 480 VAC transformation typically achieves about 98.5%-99.2%; extra North American stages (480 V to 277 V) add roughly 1-2% per stage.
5. Double-conversion UPSs attain about 94-96% at rated load and line-interactive 97-98%, but data centers may operate at only 30-50% of rated capacity.
6. In the PSU case study a roughly 180 W GPU step was followed by the PSU AC input current lagging about 1-2 AC cycles (20-40 ms), up to about 50 ms.
7. Rack PSU bandwidth of 30-50 ms bounds the fastest net swing upstream.
8. Additional buffers recommended: supercapacitors, high-slew UPS, staggered scheduling.

Using it in design

• Include rack-level templates for AI racks (32-34 kW air-cooled, 73-132 kW liquid-cooled NVL) with per-GPU power.
• Compute UPS efficiency at expected part load (30-50%) rather than rated load.
• In transient analysis treat rack PSU/power-shelf hold-up as a roughly 20-50 ms filter; swings slower than that pass to UPS and transformers.
• For large synchronous clusters, recommend supercapacitor or high-slew UPS buffering and staggered checkpoint scheduling.
• Show per-stage efficiency for AC versus 400/800 VDC chains with regional voltage variants.

Who should read it

Electrical designers, UPS engineers and power electronics specialists studying how AI loads affect the power chain.

Publication details

Authors

Yuzhuo Li; Yunwei Li

Publisher

arXiv preprint (University of Alberta)

Year

2025

Document number

arXiv:2502.01647

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.48550/arXiv.2502.01647

Licence

arXiv distribution permission — reuse conditions remain with the rights holder

Citation