Power distribution efficiency

Every kilowatt-hour that reaches a server has passed through a chain of transformers, switchgear, uninterruptible power supplies, distribution units and the server's own power supply, and each stage converts a little of the energy to heat. In an older facility, losses in the electrical chain alone can reach 10 % or more of the IT load, and the heat they produce must then be removed by the cooling system, adding further energy. Modern designs cut these losses substantially. This lesson follows energy through the power chain, explains where losses occur and why they are worse at light load, examines UPS efficiency and operating modes, distribution voltages and server power supplies, and quantifies the savings from design and operating choices.

The power chain

A typical path from the utility to the processor includes:

• Medium-voltage switchgear and step-down transformers, converting utility voltage to the building's distribution voltage.
• Low-voltage switchgear and distribution boards.
• The UPS, which in double-conversion mode rectifies AC to DC (charging batteries) and inverts it back to clean AC.
• Power distribution units (PDUs), often with transformers stepping voltage down further for the racks, and rack PDUs.
• The server power supply unit (PSU), converting AC to the DC voltages used inside the server, followed by voltage regulators on the motherboard.

Each conversion has an efficiency below 100 %. Multiplying them gives the efficiency of the chain: a chain of a 98.5 % transformer, a 95 % UPS and a 98.5 % PDU delivers about 92 % of the input energy to the IT equipment, and the rest becomes heat.

Why light load is inefficient

Most electrical equipment has two kinds of losses. Fixed (no-load) losses, such as transformer core magnetization and the control electronics and fans in a UPS, occur whatever the load. Variable (load) losses, such as resistive heating in windings and conductors, grow roughly with the square of the current. As a result, efficiency is poor at light load, where fixed losses dominate, peaks somewhere between about 40 % and 75 % of rating, and falls slightly at full load.

Redundancy makes light load the norm. In a 2N design, each of two UPS systems must be able to carry the whole load, so each normally runs below 50 %, and in a new or partly occupied facility, perhaps at 20 or 30 %. A design that is efficient at full load may therefore operate for years in its least efficient region. Modular UPS systems that switch off unneeded modules, right-sized transformers and phased build-outs help keep equipment in its efficient range.

import numpy as np

# UPS efficiency curve (double conversion) by per-unit load, interpolated from datasheet points
pts_load = [0.10, 0.25, 0.50, 0.75, 1.00]
pts_eff = [0.880, 0.940, 0.960, 0.965, 0.962]

def ups_eff(load):
    return np.interp(load, pts_load, pts_eff)

def chain_loss(it_kW, ups_kW_each, n_ups, transformer_eff=0.985, pdu_eff=0.985, eco=False):
    load_each = it_kW / pdu_eff / (ups_kW_each * n_ups)
    eff_ups = 0.99 if eco else ups_eff(load_each)
    supply = it_kW / pdu_eff / eff_ups / transformer_eff
    return supply - it_kW, load_each, eff_ups

it_kW = 1200
for label, n, eco in [("2N, two 2000 kW systems", 2, False), ("N+1, three 1000 kW modules", 3, False),
                      ("2N with eco-mode", 2, True)]:
    size = 2000 if n == 2 else 1000
    loss, load, eff = chain_loss(it_kW, size, n, eco=eco)
    print(f"{label:<30} UPS load {load:5.0%}, UPS eff {eff:.3f}, chain loss {loss:6.1f} kW, "
          f"{loss * 8760 / 1000:,.0f} MWh/yr")

2N, two 2000 kW systems        UPS load   30%, UPS eff 0.944, chain loss  109.7 kW, 961 MWh/yr
N+1, three 1000 kW modules     UPS load   41%, UPS eff 0.952, chain loss   98.5 kW, 863 MWh/yr
2N with eco-mode               UPS load   30%, UPS eff 0.990, chain loss   49.3 kW, 432 MWh/yr

In the 2N arrangement, each UPS runs at 30 % load with an efficiency of about 94 %, and the chain loses about 110 kW, nearly 1,000 MWh a year. An N+1 arrangement of smaller modules runs each at about 41 % and saves around 100 MWh a year. Operating the 2N system in eco-mode (described below) roughly halves the chain loss, saving over 500 MWh a year, before counting the cooling energy that no longer has to remove that heat.

UPS operating modes

Double-conversion (online) UPS operation gives the best power quality, isolating the load from every disturbance on the supply, but it converts all energy twice. Modern UPS systems offer higher-efficiency modes:

• Eco-mode (or bypass mode): the load runs from the utility through the static bypass, with the inverter on standby; on a disturbance, the UPS transfers to inverter within a few milliseconds. Efficiency is about 99 %, but the load is exposed to minor supply disturbances and depends on fast transfer.
• Advanced or active eco-modes: the inverter runs in parallel with the bypass, filtering harmonics and correcting power factor while most power flows through the bypass, with efficiencies around 98.5 to 99 % and faster response.
• Module management: in modular UPS systems, idle modules are switched off so that the remaining ones run at higher, more efficient load, while keeping enough capacity for redundancy.

Operators have historically been cautious about eco-modes because of the small risk to the IT load. Today, many hyperscale operators use them routinely, and server power supplies tolerate brief transfers well. The decision should be based on the site's power quality, testing of transfer behavior during commissioning (DCE-450, Lesson 5), and the customer's risk tolerance.

Transformers and distribution voltage

Transformers are highly efficient, often 98 to 99 % at moderate load, but they run continuously, so small efficiency differences add up. Codes and regulations set minimum efficiencies for distribution transformers.

In the United States, the Department of Energy sets minimum efficiencies for distribution transformers under 10 CFR Part 431, with updated standards adopted in 2024 for future production. US facilities often distribute at 480 V and step down to 208Y/120 V for IT loads, a stage that designers can avoid by distributing 415Y/240 V directly to racks, since most server power supplies accept 240 V.

Higher distribution voltages reduce current and therefore resistive losses in cables and busway, and eliminating a transformer stage removes its losses entirely. Many modern data centers distribute at 415Y/240 V to the rack, and some large operators use 48 V DC distribution within racks, converting once from AC to DC at rack-level power shelves. Direct current distribution at higher voltages (for example 380 V DC) has been demonstrated and can remove conversion stages, but it requires specialized equipment and protection.

Server power supplies

Every server has its own power supply, and their combined losses can rival the UPS. Efficiency is rated by programs such as 80 PLUS, whose Titanium level requires about 94 to 96 % efficiency across a range of loads (and 90 % at 10 % load) for 230 V internal-redundant supplies. Like UPS systems, PSUs are least efficient at light load, and servers with redundant power supplies sharing the load run each supply at half load or less. Configuring redundant supplies so that one carries the load while the other stands by, where the server supports it, and choosing supplies sized for the actual server configuration rather than the maximum, improve efficiency. Higher input voltage (240 V rather than 120 V) also improves PSU efficiency by about one percentage point.

Other electrical overheads

Several smaller loads add up: generator block heaters that keep diesel engines warm for fast starts (often several kilowatts per generator, running continuously), battery room heating and cooling, lighting, security systems, and office space within the facility. Each deserves attention: block heaters can be thermostatically controlled and insulated, lighting can use LEDs with occupancy controls, and battery technologies with wider temperature tolerance, such as lithium-ion, reduce the need for tightly controlled battery rooms.

Measuring electrical losses

Losses cannot be managed unless they are measured. Modern UPS systems, transformers with monitoring, and power meters at switchboards and PDUs allow the electrical chain to be broken down: comparing energy into and out of each stage gives its actual operating efficiency, which often differs from datasheet values because of real load levels, harmonics and temperature. An electrical pPUE (total electrical input divided by IT energy, counting only electrical losses) for each hall is a useful operational metric. Metering also reveals unexpected loads, such as heaters left on in battery rooms or unused PDUs drawing no-load losses, which are easy to fix once seen.

Batteries and energy storage

Energy storage in the UPS also affects efficiency. Valve-regulated lead-acid batteries need cool, tightly controlled rooms (around 20 to 25 °C) to reach their rated life, which means dedicated cooling energy, and they must be replaced every few years. Lithium-ion batteries tolerate warmer temperatures, last longer and occupy less space, reducing both cooling energy and replacement materials, though they require careful fire risk management. Larger battery systems can also provide grid services and support clean energy matching (Lesson 8), turning a backup asset into an active part of the energy system. Battery charging and float losses are small but continuous and should be included in electrical loss estimates.

Finally, power quality affects efficiency: harmonic currents from IT power supplies increase transformer and cable heating, and poor power factor raises currents throughout the chain. Modern IT power supplies with active power factor correction, and UPS systems that present a clean input to the grid, keep these losses small.

Taken together, these measures can bring the electrical chain's losses from 10 % or more of IT load in older facilities to around 4 to 6 % in modern ones, a saving that also reduces the cooling plant's work.