Server power supplies have moved from 500 W silver-box units to 3–5.5 kW titanium-class modules feeding 48 V or 12 V rails in dense racks. The switching stage did not change much; what changed is the loss budget. At 96% peak efficiency, every milliwatt of magnetic loss is visible on the efficiency curve, and the thermal headroom inside a 1U chassis is close to zero.
1. What the topology demands from the magnetics
A typical titanium server PSU runs a bridgeless or interleaved totem-pole PFC front end, followed by an LLC or phase-shifted full bridge, then a synchronous rectifier output stage. Each block has a different magnetic requirement:
| Stage | Magnetic | Dominant requirement |
|---|---|---|
| PFC | Boost / interleaved boost inductor | Low core loss at 50–100 kHz with large DC bias, high saturation headroom |
| DC-DC | LLC resonant inductor + transformer | Controlled leakage inductance, low proximity loss, tight coupling |
| Output | Output inductor or integrated planar output | Very low DCR at 100–200 A, low height |
| EMI | Common mode choke + differential choke | High impedance at 150 kHz–30 MHz in a small footprint |
2. PFC inductor: powder core still wins
Interleaved PFC at 3 kW per phase carries 15–25 A RMS with a 60–120 µH inductance. Distributed-gap powder cores (sendust, high flux, iron powder) hold inductance under bias far better than gapped ferrite, and their soft saturation prevents the catastrophic current spike a gapped ferrite core produces at the knee. Select for loss at the ripple frequency, not at 60 Hz, and verify the bias curve at the actual peak current — the 20% inductance drop point is where the ripple and the EMI spectrum change.
3. LLC transformer: leakage inductance is a design parameter
In an LLC the resonant inductance is often realised as the transformer’s own leakage. That makes winding geometry part of the control loop: a change in bobbin, interleaving or insulation thickness shifts the resonant frequency and the gain curve. Practical rules:
- Fix the winding sequence in the drawing and treat it as a controlled characteristic.
- Specify leakage inductance with a tolerance (±10% is realistic for a wound part) and measure it at the same frequency the design assumes.
- Use planar or foil windings where height allows; the reduced proximity loss is worth more than the tooling cost at this power level.
- Watch the hot spot at the winding-to-core interface — it is the limiting point in a fan-cooled 1U chassis.
4. Loss budget, in numbers
For a 3 kW unit at 96% efficiency the total loss is 125 W. A realistic split for the magnetics: PFC inductor 8–12 W, LLC transformer 10–16 W, resonant inductor 3–6 W, output inductor 6–10 W, EMI chokes 2–4 W. That is roughly 20% of the entire loss budget, which is why core grade selection matters more here than in a 60 W adapter.
5. Thermal and mechanical reality
Airflow in a 1U chassis is 20–40 CFM across the whole unit, and the magnetics sit downstream of the switching devices. Design to a 40 K rise at 55 °C inlet, not to a 25 °C bench. Potting or varnish improves coupling to the air but adds 3–5 K of its own thermal resistance if the compound is not void-free.
6. Reliability items that matter in a data hall
Hot-plug inrush, 40 °C–65 °C storage, 8–12 year service life, and fan-failure ride-through at reduced load. Inrush and fan failure are both magnetic stress events: specify the saturation headroom for the worst case, and qualify the insulation for the highest hot spot the unit can reach with one fan dead.
Shaanxi SiRui Technology supplies the Magnetic Core series for exactly this application — distributed-gap powder cores for interleaved PFC, low-loss Mn-Zn ferrite for LLC transformers, and nanocrystalline cores where a compact high-impedance common mode choke is needed.
Send your PSU specification to sales@sirui-tech.com for a loss-budget-based core proposal.