An introduction to three-phase power and PDUs

Why Three-Phase Matters in the Rack

Every server, switch, and storage device in a data center depends on a power distribution unit (PDU) to receive electricity. These industrial-grade power strips are typically rated to supply the equivalent of several households’ worth of power. Choosing the right PDU depends on the local grid’s voltage, phase, and plug type, and getting that choice wrong can waste money or risk downtime.

In a standard rack setup, each server with dual power supplies connects to two separate PDUs—one via a red cable, the other via a blue cable. This provides redundancy: if one PDU or one power supply fails, the other feed keeps the server running.

The Basics: AC Power and Its Math

PDUs and servers run on alternating current (AC), where voltage and current oscillate as sine waves between positive and negative values. A 100V feed, for example, swings between +100V and -100V at a set frequency. One complete cycle is a phase of 360 degrees; at 50Hz, this happens 50 times per second. This alternating behavior follows Faraday’s Law, which states that a changing magnetic field induces current. In a simple generator, a rotating magnet induces a fluctuating current in a nearby coil—current rises as the magnet’s north pole approaches, drops to zero as it passes, then reverses direction as the south pole comes near.

Because current and voltage fluctuate in tandem, their product (power) is always positive, but it is not constant. To get a steady value, engineers use root-mean-square (RMS) averaging, which divides the peak by √2. When a spec sheet says 208V, that is the RMS value; the actual voltage peaks at 208V × √2. So a PDU with a 5kW RMS capacity is actually handling peaks of about 7.1kW.

Load calculation is straightforward: sum the power draw of all machines in a rack and keep it below the PDU’s rated capacity. For a typical US single-phase PDU at 208V and 24A, that cap is 5kW. If demand exceeds that, you need more PDUs or racks—potentially leaving hardware and floor space underutilized.

The Three-Phase Advantage

A three-phase PDU offers up to √3 (about 1.7) times the capacity of a single-phase unit. The same 208V/24A specifications yield an 8.6kW PDU on three phases. While three-phase equipment costs more and has heavier cabling, it can consolidate what would otherwise require two single-phase racks. For the same total hardware load, fewer PDUs, fewer racks, and fewer power feeds are needed. That trims both capital and operating expenses and reduces the physical footprint.

Three-phase units also carry an extra circuit breaker compared to single-phase models. A 48-outlet single-phase PDU splits into two circuits of 24; a three-phase unit divides into three circuits of 16. If a breaker trips, fewer outlets are lost—16 instead of 24—so the blast radius of a fault is smaller.

How Three Phases Are Generated and Wired

A three-phase generator contains three coils spaced 120 degrees apart inside the rotor. As the magnet spins, each coil produces its own single-phase waveform, offset by a third of a cycle. This yields three live conductors—L1, L2, and L3—each carrying its own phase voltage and current. Any pair of phases forms a line with its own voltage and current.

Terminology can be confusing: line voltage and phase voltage are identical in single-phase systems, but in three-phase circuits, the relationship depends on topology. There are two wiring methods:

• Delta: V_line = V_phase; I_line = √3 × I_phase
• Wye: V_line = √3 × V_phase; I_line = I_phase

Wye circuits also include a neutral line where all three phases meet, enabling a second, lower voltage for small devices. Common Wye PDU ratings in North America are 120V/208V or 230V/400V. Notably, the generator and the PDU need not use the same circuit type.

Why the √3 Multiplier Exists

The √3 factor comes from adding phasors—complex numbers representing sine waves. Adding two phases L1 and L2 is like vector addition: walk one mile east, then one mile north; you’ve moved two miles but only traveled 1.4 miles from your start. For two identical coils offset by 120 degrees, the magnitude of the summed vector is √3 times each individual phase magnitude. Since voltage is a scalar quantity for power calculations, only that magnitude matters.

With all three line powers equal, total effective power is three times a single line’s power. Simplifying the algebra across both Delta and Wye configurations yields the key result:

P_overall = √3 × V_phase × I_phase

For a US PDU at 208V and 24A, that computes to 8,646W or roughly 8.6kW—again, 1.7 times the single-phase capacity.

The Real-World Problem: Phase Imbalance

The √3 advantage assumes perfectly equal loads on all three phases. In practice, servers vary in power draw, and they are not always distributed evenly across the PDU’s branches. Imbalance directly reduces usable capacity. A balanced 8.6kW PDU spreads 2.88kW per branch. If one branch loses 1kW of load and another gains it, the PDU can exceed its limit even when the total is still 8.6kW—because phase currents are what matter, not just total power.

Each phase is limited to 24A. Overloading one branch trips the breaker, which is typically rated lower (e.g., 20A) for protection. Loading a single branch to its limit effectively turns a three-phase PDU into a single-phase one, forfeiting the √3 benefit entirely.

To avoid these issues:

• Keep total rack power under the PDU’s maximum.
• Spread cabling as evenly as possible across all three branches.
• Verify that each branch’s phase current stays below the fuse or breaker rating.

Spreadsheets that calculate current per phase on a specific PDU model are useful during rack design.

Practical Example: Dense vs. Fine-Grained Servers

Consider a rack with four 2,600W-class 2U4N chassis—each drawing 1,800W. Four chassis total 7.2kW, which fits under the 8.6kW PDU cap. Adding a fifth would hit 9kW and is not possible. But splitting four chassis across three branches inevitably leaves two on one branch. Checking phase currents, that branch may only have about 1.1A of headroom before hitting the 24A limit—leaving room for just a few hundred watts. Adding that load on the wrong branch trips the breaker; putting it on another branch helps balance.

Smaller, more efficient servers offer an alternative. If the same physical space and power budget can instead host 18 nodes at 450W each—the functional equivalent of splitting each 2U4N into four independent, power-supply-equipped 1U servers—then more compute fits under the same 8.6kW cap. The finer granularity means load can be placed more precisely across the three phases, yielding better balance and an additional 12.5% more nodes without changing the rack’s power provisioning.