Skip to content
Wattality
ENTR
Technology

Technology · Power distribution

Power distribution in the AI data center

Distribution stages from the grid connection to the GPU rack, redundancy architectures and the dynamics of training load. Wattality's products are designed to work at specific stages of this chain.

Distribution stages

From the grid connection to the GPU rack

Drawn topology: 2N distribution, A side and B side. Cooling loads on a side branch from side B.

  1. 01Stage

    Grid connection

    Function
    Point of interconnection and main transformers. Two independent medium-voltage (MV) feeds.
    Risk
    Fast load changes appear at the point of connection. Interconnection terms can limit ramp rate and oscillation.
    Data gap
    Meters record averages. Second-level load and IT activity are not in one record.
    Wattality
    Site is designed to run grid, generators and batteries as one system.

  2. 02Stage

    MV switchgear, A and B sides

    Function
    Each side is sized for the full critical load. Generators connect at the MV bus.
    Risk
    A shared bus, breaker, cable route or room ties both sides to one failure.
    Data gap
    Relay event records stay in the devices. Post-event sequencing is done by hand.
    Wattality
    Site is designed to track breaker states and generator readiness on one timeline, from outside the protection chain.

  3. 03Stage

    MV uninterruptible power supply (UPS) and battery

    UPS type

    Function
    Double conversion: the load is fed from the inverter, the battery sits on the DC link. Line-interactive: the load is fed from the source, the battery sits on a shunt converter.
    Risk
    End-of-life capacity is assumed with a fixed aging factor. Power capability is not measured in service.
    Data gap
    Cell internal resistance and chemical state. Monitoring usually stops at block or module voltage and temperature.
    Wattality
    Sense is being developed for cell-level measurement, Core for pack control and DC/DC for energy transfer between packs.

  4. 04Stage

    MV/LV transformer

    Function
    Steps medium voltage down to the hall's low voltage (LV).
    Risk
    Harmonics and load cycling raise winding temperature. Winding temperature sets insulation life.
    Data gap
    Winding temperature stays in the transformer relay and is not logged on the load profile's clock.
    Wattality
    Site is designed to log transformer data on the same clock as the load profile.

  5. 05Stage

    LV switchboard

    Function
    Hall and rack feeders. Protective device coordination between breakers.
    Risk
    If coordination is lost, a fault moves to the upstream breaker and more racks lose power.
    Data gap
    Feeder loads and trip records sit in power monitoring, on a separate clock.
    Wattality
    Site is designed to merge feeder data with UPS and battery data in one data model.

  6. 06Stage

    Server rack: busway, PDU, PSU, BBU

    Function
    A and B busways, power distribution units (PDU), dual-corded server power supplies (PSU), in-rack battery backup unit (BBU).
    Risk
    GPU load moves in synchronized steps. The swing propagates up the chain.
    Data gap
    Power inside the rack changes within milliseconds. Upstream systems see averages over seconds.
    Wattality
    BBU is designed to meet fast load steps on the rack power line.

  7. 07Stage

    Cooling and mechanical loads

    Function
    Chillers, pumps and coolant distribution units (CDU). Separate mechanical branch.
    Risk
    With liquid cooling, a stopped pump heats the GPUs quickly. Some pumps need uninterruptible supply.
    Data gap
    Mechanical power and IT load are not on one timeline.
    Wattality
    Site is designed to plan the order in which mechanical loads return after a source transfer.

Load dynamics

UPS type and the in-rack BBU set how much of a load step reaches the source.

In a training job the GPUs compute and wait in step. Hall load shows a start-up ramp, synchronized steps, checkpoint dips and abrupt stops.

BBU is designed to meet fast steps. Site is being developed to assign each change to a source.

Training load, layer by layer
Schematic

UPS type

BBU at the rack

The swing travels up from below

  1. Generator or grid

    Slow ramp

  2. UPS and its battery

    Input ramp limited, difference from the battery

  3. BBU in the rack

    Takes the fastest steps

  4. GPU load

    Steps, dips, oscillation

The BBU meets the fastest steps inside the rack. The UPS limits how fast its input power changes and covers the difference from its battery. The source sees a slow ramp.

Load passed upLoad arriving from belowCarried by the batteryShapes only, no scale

Redundancy architectures

N+1, 2N, distributed redundant and block redundant distribution

  • Two sides are independent when they share no bus, breaker, cable route or room.
  • A single-corded load hangs on one side; it reaches the second side only through a static transfer switch (STS).
  • Site is designed to track whether each UPS and battery can carry the load it takes on after a single failure.
Redundancy architecture
Schematic

Architecture

State

2N, A and B sides

A sideB sideSourceUPSPDUPSU APSU BServer rack

The two sides share the load. The PSUs draw from both feeds.

Normal loading
≤ 50% per side.
Single failure
One side lost: dual-corded PSUs draw from the other side. No transfer.
Concurrent maintainability
A whole side can be taken out for maintenance.
Batteries
Each UPS has its own battery. The sides are separate.
NormalLost

Facility data

Four control systems, four separate clocks

The battery management system (BMS), the UPS controller, the electrical power monitoring system (EPMS) and the generator controls each log on their own clock and in their own format. Post-event sequencing is done by hand.

Facility control systems
Schematic

Site · supervisory layer

Setpoints and operating modes

  • Battery management (BMS)

    Cells and modules

    Own clock
    Own data format

  • UPS controller

    Rectifier, inverter, battery

    Own clock
    Own data format

  • Power monitoring (EPMS)

    Meters, breakers, feeders

    Own clock
    Own data format

  • Generator controls

    Engine and regulators

    Own clock
    Own data format

Protection · trips and interlocks hardwired

Each system logs on its own clock. The same event appears at four different times in four records.

Site: one clock, one data model, outside the protection chain

Designed as a supervisory layer for setpoints and operating modes. It takes no part in trip decisions. Trips and interlocks stay hardwired.

Fleet is a separate cloud product that connects sites and develops the models.

Design

We design the power distribution architecture and redundancy structure for AI data centers and critical facilities.

  1. 01Power distribution architecture and redundancy topology (N+1, 2N, distributed redundant, block redundant)
  2. 02UPS and battery sizing
  3. 03Generator step-load and load-dynamics studies
  4. 04Short-circuit studies and protective device coordination
  5. 05Single-line diagrams
  6. 06Battery and facility control layer integration
  7. 07Commissioning support
Write to us about a design

Continue

Let's talk about your facility's power distribution.

Write to us about redundancy architecture, UPS and battery sizing or facility data.

Get in touchinfo@wattality.com