Products · Solutions

GPU Server Power Supply Solutions

Accelerators rewrote the power budget of the rack. A single H100 SXM draws 700W, a B200 draws 1000W, and an 8-GPU HGX node crosses 10kW of system load before storage and networking say a word. This page lays out how TitanWatt production parts — titanium CRPS arrays today, 54V M-CRPS in development — combine into power architectures per node type, with the telemetry to see trouble coming.

What GPUs Demand From the Power Chain

Reference Point	Figure	Power-Chain Consequence
NVIDIA H100 SXM GPU	700W per GPU	Board VRMs need stiff 12V/54V feed, low ripple under transient load steps
NVIDIA B200 GPU	1000W per GPU	Per-GPU draw now exceeds one 1300W PSU's comfortable share — arrays mandatory
12VHPWR card connector	600W ceiling	Card-level cabling margin vanishes; distribution voltage must rise
HGX 8-GPU node	5.6kW–8kW bare GPU, 10kW+ system	3–5.5kW PSUs × N+1, no single-module failures tolerated
AI rack density	40–120kW per rack	Shelf-level 54V distribution; traditional 5–10kW rack assumptions obsolete
Load transients	Sub-millisecond full-throttle steps	Hold-up and sharing control must react faster than host throttling
Fleet duty cycle	Near-continuous training load	Efficiency at 50–80% load dominates TCO; titanium 96% @ 50% matters

Architecture

Power Topology, Stage by Stage

STAGE 1

AC Input

200–240Vac branch circuits land on rack PDU outlets; C19 feeds per PSU. Power factor ≥ 0.95 at 50% load keeps the branch honest.

→

STAGE 2

CRPS Array

3–6 hot-plug titanium modules, 1600W–2400W each today, active current sharing across the bus; N+1 with PMBus health per slot.

→

STAGE 3

54V Busbar

54V (or 12V for legacy nodes) distribution: the same power through a fraction of the current, keeping line loss and connector heating manageable at multi-kilowatt draw.

→

STAGE 4

GPU Boards

On-board VRMs step down to sub-volt rails per GPU. The PSU's job ends at a clean, telemetry-visible feed and fast throttling signals.

Today: stages 1–2 built from in-production 1600W–2400W CRPS titanium. In development: stages 2–3 native at 54V via the M-CRPS 3000W–5500W program.

Sizing

Solutions by Node Type

In production

1–2U GPU Servers

One to four low-profile accelerators, 1.5–3kW system draw. Two to three 1300W–1600W CRPS units in N+1 at 12V — no intermediate conversion stage, minimum parts count, full PMBus visibility.

In production

HGX-Class Nodes

8-GPU nodes at 10kW+ system load. Four to five 2400W CRPS front-end units with active current sharing, oversubscribed 5×2.4kW = 12kW against a 10kW draw, or N+N cabling across two PDUs for dual-domain fleets.

Development Samples

Full AI Racks

40–120kW racks want shelf-level 54V. Our 5.5kW M-CRPS development samples target the ORv3 slot so the same module serves nodes now and shelves next — one qualification across your roadmap.

CRPS power supplies feeding GPU compute boards inside a dense AI server node

Redundancy & Telemetry

See the Failure Before It Costs $9,000

Uptime Institute pegs data center downtime near $9,000 per hour, and in a GPU cluster the power front-end is the single most exposed shared element. Every TitanWatt smart PSU streams voltage, current, temperature and fan RPM over PMBus 1.2, so your BMC can trend capacitor aging and fan bearing wear instead of discovering them mid-training run.

When a module does drop, power throttling lets the host shed load in an ordered sequence, and cold redundancy brings a rested spare online that has already proven itself. The failure becomes a maintenance ticket scheduled between training jobs — not a terminated checkpoint.

PMBus telemetry details

Two CRPS power supplies installed at the rear of a GPU compute node

Inside the Node

Where the Power Front-End Lives

In an 8-GPU node the supplies sit at the very back of the thermal path — last in the airflow, first to be blamed. Handle-out access, front or back airflow options and per-rail telemetry exist so that the most exposed shared component is also the easiest to inspect and replace.

Field Record

From Our Project Log

AI cluster front-end program: a customer building a training cluster engaged us on the 5.5kW M-CRPS path while standardizing our in-production 2000W CRPS units for their existing 2U nodes — one supplier, both rungs of their ladder, with the 54V transition designed in rather than bolted on. Project details are anonymized in our projects log.

FAQ

Should my GPU node run 12V or 54V?

Below roughly 3kW per node, 12V CRPS arrays are simpler and cheaper. Above that, current in the 12V path makes copper and losses grow fast — 54V shrinks both the current and the copper losses dramatically. The crossover is why our 1600W–2400W series serves 2U–4U nodes while the M-CRPS 54V program targets shelf-scale racks.

How many PSUs does an 8-GPU HGX node need?

Budget the node at 10kW or more of system load. Five 2400W units in a 4+1 array give 9.6kW active plus a full spare; six in 5+1 give 12kW active with margin for transient peaks. We validate the sharing balance at your exact unit count during first-article testing.

Do your PSUs support power throttling signals for GPUs?

Yes. On PMBus variants the throttle signal and health telemetry integrate with standard BMC implementations: when redundancy degrades, the host can clock GPU workloads down before the array loses margin, keeping the job alive on reduced power instead of losing it entirely.

What is available today versus in development?

Today: 550W–2400W titanium CRPS in production, samples in 2–4 weeks, volume in 6–10 weeks. In development: 3000W–5500W M-CRPS as development samples with target specs, for joint evaluation. We label the difference on every page rather than blurring it.

Size Your GPU Power Chain

Send your node or rack configuration — we reply within 48 hours with an array sizing and sourcing plan.