MiDEN Technical Insights

High Frequency Transformer for AI Server PSU

Introduction

High frequency transformer for AI server PSU selection is now a power density problem as much as an electrical one. Rack power supplies have moved from a few hundred watts to multi-kilowatt units delivering a 48 V bus, and the transformer in the isolated DC-DC stage largely sets the height, the loss and the thermal budget of the whole unit. The sequence below is the order the constraints actually arrive in: switching frequency and flux swing first, then core and winding construction, then leakage inductance, isolation and thermal data.

Technical Analysis: Frequency Sets the Flux Swing and the Window

Once the topology and the input range are fixed, the switching frequency sets the flux swing and the flux swing sets the core. Faraday's law ties the three together: the primary winding sees a volt-second product proportional to Vin × D / fsw, and dividing that by the primary turns and the effective core area gives the flux the core has to absorb. Raise the frequency and the flux swing falls in proportion, so a smaller core area can carry the same power. That is the whole reason server supplies moved from tens of kilohertz to hundreds of kilohertz, and it is also where the trade begins.

Core loss does not fall with the flux swing alone. Ferrite loss rises steeply with frequency, so at a fixed flux swing a doubling of frequency more than doubles the loss per unit volume. The curve that matters is the one on the material datasheet at the working temperature: a material that looks excellent at 100 kHz and 25 °C can be a poor choice at 300 kHz and 100 °C.

Frequency also moves loss out of the core and into the winding. Skin depth shrinks, the proximity effect grows, and the AC resistance of a round wire winding can reach several times its DC value. This is the point at which construction stops being a mechanical detail: foil, Litz and planar windings all exist to keep current spread across the copper instead of pushed to its edges.

For the isolated stage of an AI server power supply the practical window today is 150 kHz to 400 kHz. Below it the transformer is too tall for a 1U chassis; above it core loss, winding AC loss and the switching loss of the primary devices all rise at once.

Design Challenges in a Server PSU Transformer

Five constraints decide whether the part works in a rack rather than on a bench.

• Height. A 1U chassis leaves roughly 40 mm of internal height and the transformer has to sit inside the airflow path. That budget usually decides planar or foil construction before any electrical argument does.
• Thermal margin. The transformer sits in the same airstream as the semiconductors and downstream of them, so it sees a raised inlet temperature. Fix the ambient and the permitted rise, for example 50 °C inlet with a 40 K rise, and design the losses to that point rather than to 25 °C.
• Leakage inductance and interwinding capacitance. Higher frequency makes both matter more: leakage inductance pushes loss into the primary snubber, while interwinding capacitance drives common mode current across the isolation barrier. Sectionalised or interleaved windings trade one against the other.
• Isolation coordination. The barrier separates the mains-derived primary from the 48 V secondary, so creepage, clearance and the insulation system are safety parameters. They belong in the drawing package, not in a follow-up email after the sample arrives.
• Production consistency. Turns count, layer placement and winding tension all shift leakage inductance and loss. A design that depends on one hand-wound sample is not a design that ships at volume.

Application Scenarios: 3 kW to 5.5 kW Rack Supplies

The same selection sequence lands in different places depending on the power level and the distribution voltage.

• 3 kW, 1U, 48 V output. The common density point. A planar transformer with few turns and a modest core area is usually the only construction that fits the height, and the primary devices carry most of the loss.
• 5.5 kW and above. Current rather than voltage becomes the problem. Foil or heavy planar windings carry the secondary current with acceptable AC loss, and the mechanical design has to move heat out of the winding stack instead of relying on the core.
• 400 V DC distribution. Distributing at a higher DC voltage changes the primary side of the isolation stage and eases the secondary current problem, but the barrier still has to be coordinated for the same working voltage.

In all three cases the design review comes back to the same four numbers: turns ratio, flux swing, leakage inductance and the thermal condition.

Product Connection: Which MiDEN Family Fits

The transformer position is served by three families, and which one applies follows from the height budget and the isolation drawing rather than from power level alone.

• The transformer series covers wound constructions for the lower end of the frequency window, where a conventional bobbin still fits the chassis.
• The planar transformer series covers the low profile case: a 1U server supply running between 150 kHz and 400 kHz, where layer control also keeps leakage inductance repeatable.
• Custom magnetic components are the route when the isolation system, the creepage distance or the mounting envelope rules out a standard body.

MiDEN builds high frequency transformers to a customer drawing as well as to standard bodies, so a selection enquiry can start from those four numbers instead of from a catalogue part number. The method behind them is set out in the high frequency transformer design guide, and the wider component set is listed under magnetic component applications.

Frequently Asked Questions

What switching frequency should an AI server PSU transformer run at?
Most rack supplies sit between 150 kHz and 400 kHz. Below that the transformer is too tall for a 1U chassis; above it core loss, winding AC loss and primary switching loss all rise together.

Planar or wound construction?
Planar wins when height and repeatability dominate, which is the usual case at 1U. A wound construction stays practical below roughly 150 kHz, or when the isolation drawing needs a creepage path a planar stack cannot provide.

How is the isolation requirement specified?
As a working voltage plus creepage and clearance distances and an insulation class, not as a single voltage number. The barrier is a safety parameter, so the requirement belongs in the drawing package and the RFQ.