800 VDC Architecture for Next-Generation AI Infrastructure

Overview

The paper opens with the step changes in rack power that come from growing NVLink domains. Moving from Hopper to GB300, TDP rose 75%, performance about 50x and rack power density 3.4x, because four separate 8-GPU NVLink domains became one 72-GPU domain. The authors conclude that this trajectory cannot be sustained on 415/480 VAC distribution and that megawatt-class racks need a different distribution voltage.

The second theme is load dynamics. Synchronised LLM training swings rack power from roughly 30% at idle to 100%, with transients as fast as about 400 microseconds, idle gaps up to 100 ms, checkpointing pauses of 1-5 s and ramps over minutes. A 50% duty square wave whose peak is 50% above the average raises RMS losses by 25%, so upstream equipment has to be sized for both peak and RMS effects. The proposed remedy is capacitors near the racks and a BESS at the utility/generator interface.

The comparison table centres on conductor utilisation: power per mm2 of conductor is 0.6 kW at 415 VAC, 0.8 kW at 480 VAC (+16%), 1.7 kW at 800 VDC (+157%) and 3.1 kW at 1500 VDC (+382%). Typical AC whips are capped at 60 A or 100 A (IEC 60309), which constrains AC delivery to MW-class racks. The paper also reviews rival DC schemes (750 VDC, +/-400 VDC, 1500 VDC) and explains why it picked single-ended 800 VDC: widely available two-pole breakers and no pole load balancing.

On the architecture side, the 800 VDC MGX rack steps directly from 800 V to 12 V with a 64:1 LLC converter, saving about 1% efficiency and 26% area versus multistage conversion. The migration path runs from side power racks to facility rectifiers and then MV rectifiers or solid-state transformers: LV rectifier strings target about 1.5 MVA at roughly 99% efficiency, MV units up to 7.5 MVA at 98.5% or better, for example converting 35 kV AC to 800 VDC. It is a vendor paper, and timelines depend on NVIDIA's own roadmap.

Key points

1. Hopper to GB300: TDP +75%, performance ~50x, rack power density 3.4x.
2. Synchronised LLM load swings rack power between about 30% and 100%; transients ~400 us, idle gaps ~100 ms, checkpointing 1-5 s.
3. A 50% duty square wave with peak 50% above average raises RMS losses by 25%.
4. Power per conductor area: 415 VAC 0.6, 480 VAC 0.8, 800 VDC 1.7, 1500 VDC 3.1 kW/mm2.
5. Typical AC whips are limited to 60 A or 100 A (IEC 60309).
6. LV rectifier strings ~1.5 MVA at ~99%; MV rectifier/SST up to 7.5 MVA at >=98.5%.
7. In-rack 64:1 LLC conversion from 800 VDC to 12 V saves ~1% efficiency and 26% area.
8. Single-ended 800 VDC was chosen over +/-400 VDC; 1000 V-class parts near term, 1500 VDC a possible long-term target.

Using it in design

• Add a single-ended 800 VDC distribution option to the electrical model, with rectifier blocks of ~1.5 MVA (LV) and up to 7.5 MVA (MV/SST) and the stated efficiencies.
• Model AI racks with a dynamic profile (30-100%, millisecond scale) instead of flat kW, and size upstream gear for peak and RMS uplift.
• Warn when planned rack kW exceeds what 60 A or 100 A AC whips can deliver.
• Represent storage in two layers: rack-side capacitors and a BESS at the utility/generator boundary.
• Use the kW/mm2 ratios as a first estimate for DC conductor sizing.

Who should read it

Design engineers and decision makers choosing electrical distribution for AI campuses.

Related standards

• IEC 60309
• IEC 62477-1
• NEC Article 690
• VDE-SPEC 90037
• UL 62368-1

Publication details

Authors

Jared Huntington; Mike Tu

Publisher

NVIDIA

Year

2025

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