Why Magnetic Components Set the Efficiency Ceiling in AI Server Power Supplies

AI servers are asking power supplies to deliver more energy from less space while meeting tighter efficiency requirements. Under the official 80 PLUS certification criteria, Titanium requires 230V/277V redundant data center power supplies to reach 96% efficiency at 50% load. The newer Ruby tier raises this to 96.5%, or 97% for specified higher-voltage AC/DC input classes. As GaN and SiC reduce semiconductor switching losses, the remaining loss budget shifts toward transformer and inductor core loss, AC copper loss, leakage inductance, parasitic capacitance and thermal rise.

1. Industry Overview: The Efficiency Race in AI Data Centers

Server power supplies must deliver more power in less space—within a 1U rack unit or compact power modules. AI accelerators keep pushing per-rack power upward: the Open Rack V3 specification now includes a 12kW PSU module, while next-generation systems are moving toward higher-power three-phase shelves. Power density and heat flux rise in lockstep.

1.1 The Efficiency Bar Keeps Rising

The closer efficiency gets to its ceiling, the harder it becomes to compensate for every watt of magnetic loss elsewhere in the power train.

1.2 Higher Frequency Is the Path to Power Density

Raising switching frequency reduces the volt-seconds each cycle must handle, shrinking the required core cross-section. A transformer designed for 500kHz can be significantly smaller than its 100kHz equivalent. TI's PMP23146 server auxiliary power reference design uses a planar transformer at 400kHz, illustrating how frequency and planar magnetics can work together to compress volume.

But higher frequency also pushes core, winding, insulation, and parasitic parameters into the dominant zone. The design target is no longer "the highest frequency"—it is the system optimum across magnetic volume, semiconductor losses, passive losses, EMI, thermal rise, and cost.

2. Six Mechanisms That Make Magnetics the Constraint

2.1 Core Loss Scales with Frequency and Flux Swing Together

Higher frequency allows a smaller core cross-section, but hysteresis and eddy-current losses climb. Ferrites, nanocrystalline alloys, and metal powder cores each behave differently across frequency, temperature, DC bias, and waveform shape. Datasheet typical values cannot substitute for testing under real converter waveforms.

2.2 AC Copper Loss Is Not DC Resistance

Skin effect concentrates current at the conductor surface; proximity effect distorts current distribution in adjacent windings. At high frequency and high current, the effective resistance of copper foil, flat wire, and PCB windings rises dramatically. Layer arrangement matters more than simply adding copper thickness.

2.3 Leakage Inductance and Parasitic Capacitance Pull in Opposite Directions

Reducing primary-to-secondary spacing lowers leakage inductance but increases interwinding parasitic capacitance and common-mode noise. Increasing insulation distance improves safety but raises leakage inductance and volume. Planar transformers must simultaneously optimize layer stacking, shielding, and winding interleaving.

2.4 LLC and DAB Topologies Demand Controllable Magnetic Parameters

In resonant LLC converters, magnetizing inductance and resonant inductance set the gain curve, soft-switching range, and transient response. In dual-active-bridge (DAB) converters, leakage inductance itself becomes an energy-transfer element. Parameter variation directly impacts efficiency and current sharing across paralleled units.

2.5 High Power Density Hides Hotspots Inside the Magnetic Component

Internal thermal resistance in cores and windings is high—a normal exterior temperature does not rule out internal hotspots. Potting, clamps, busbars, and cold plates reshape the thermal path. Thermal design must be verified together with electromagnetic simulation, material lifetime, and insulation class.

2.6 Production Consistency Determines Whether Lab Designs Scale

Air-gap dimensions, winding placement, copper thickness, lamination, magnetic material batches, and assembly pressure all shift inductance and loss. Automated winding, PCB-based planar magnetics, and in-circuit testing earn their keep by converting the laboratory optimum into a stable manufacturing process window.

3. Engineering Practice: What This Means for Magnetic Component Design

For magnetics suppliers and power supply engineers, the shift is concrete:

Material selection must be waveform-verified. Loss curves measured with sinusoidal excitation understate losses under the square-wave, quasi-square, or resonant waveforms found in LLC and DAB stages. Ask for loss data under your actual operating waveform, flux swing, and temperature.

Winding design is an AC problem. Dowell-based layer analysis, interleaving, and litz-wire strand-diameter selection (e.g., 0.05–0.1mm strands sized to the skin depth at your switching frequency) should be part of the quotation package, not an afterthought.

Parasitics are design parameters, not defects. In DAB designs, leakage inductance can be deliberately engineered as the energy-transfer element; in LLC, the ratio of magnetizing to resonant inductance defines the operating region. Specify tolerances and verify them across production batches.

Thermal validation must reach the hotspot. Rely on thermocouple-embedded prototypes or thermal imaging of cross-sectioned samples—not just surface temperature—to validate winding hotspots and insulation margins.

Consistency is a deliverable. Documented process windows for air-gap assembly, winding tension, and core batch traceability are what allow a validated 96.5% design to ship at volume.

3.1 How TrafoPSU Approaches This

As a custom magnetic components manufacturer, TrafoPSU develops high-frequency transformers, planar transformers, resonant inductors and common-mode chokes for power conversion applications from 50W to multi-kilowatt. The engineering workflow follows the same system-level constraints:

Core material portfolio spanning high-frequency ferrites through nanocrystalline alloys, selected against measured loss behavior at the customer's operating frequency and temperature

Winding technologies including litz wire, flat wire, copper foil, and multilayer PCB planar structures, with AC resistance analyzed and verified at the design frequency

Topology-aware parameter control for LLC resonant tanks and DAB leakage-integrated designs, with defined production tolerances

EMI-side components—common-mode chokes engineered for target impedance bands and self-resonant frequency placement in high-frequency systems

These principles apply across planar transformers, high-frequency transformers, resonant inductors and common-mode chokes used in data center and server power applications.

FAQ

Why do magnetic components limit server power supply efficiency?

After GaN and SiC devices reduced switching losses, core losses (hysteresis and eddy currents), AC copper losses from skin and proximity effects, leakage inductance spikes, and parasitic capacitance now dominate the remaining loss budget. At 80 PLUS Titanium (96% at 50% load) and Ruby (96.5–97%) levels, every watt of magnetic loss is difficult to compensate elsewhere.

What is the 80 PLUS Ruby efficiency requirement?

For redundant data center PSUs, 80 PLUS Titanium requires 96% efficiency at 50% load for 230V/277V internal units. Ruby raises this to 96.5%, while the specified 380V/800V DC and 400V/480V AC input classes require 97% at 50% load.

Why are planar transformers preferred for high-frequency server power supplies?

Planar transformers use PCB or copper-foil windings that enable low-profile designs and automated manufacturing, with better thermal spreading and predictable parasitics than wire-wound equivalents—though they demand careful optimization of stacking, shielding, insulation, and thermal paths.

How does switching frequency affect transformer size?

Higher switching frequency reduces volt-seconds per cycle, allowing smaller core cross-sections—a 500kHz transformer can be significantly smaller than a 100kHz design. However, core losses, AC winding resistance, EMI, and thermal rise all increase with frequency, so the goal is a system optimum, not the highest frequency.

Topics: Server Power Supply | Magnetic Components | High-Frequency Transformer | Planar Transformer | LLC Resonant | DAB Converter | GaN / SiC | Core Loss | 80 PLUS Titanium | Data Center Power