Power Inductor Selection for Multiphase VRM — What Drives Core-Rail Magnetics

Why the inductor, not the controller, often decides VRM performance #

A multiphase VRM looks like a controller story: pick a PWM controller, bolt on a few smart power stages, and let the loop software balance the phases. In practice, the component that quietly governs ripple, transient response and how hot the board runs is the output inductor on each phase. The controller sets timing and the power stage sets the switch, but the energy that smooths the current into the load lives in the magnetics.

For a buyer or a design engineer specifying a core rail — whether that is a CPU socket, an AI accelerator or an FPGA — the inductor is a first-order decision. It is also one of the easiest parts to under-specify, because the headline number on a distributor page (the inductance in microhenries) is the least important thing to get right.

What a VRM inductor actually does #

Each phase of a buck converter is a current source that charges and discharges a coil. The inductance sets how much current ripple flows for a given switching frequency and duty cycle:

``` ΔI = Vout × (1 − D) / (fsw × L) ```

Smaller L gives a larger ripple; larger L gives a smaller ripple. That single relationship drives almost every trade-off below. The inductor does not regulate voltage by itself — it buffers the energy the power stage delivers so the load sees a steady current rather than a square wave.

Key parameters to read before you buy #

The two current ratings are the ones most often confused. Isat is about the magnetic material — the point where inductance drops (often specified as a 20–30% fall). Irms is about self-heating — the current that raises the part's temperature by a stated amount. A part can be within Irms yet still saturate under a load step, or pass Isat yet run too hot. Both must clear the application.

The core trade-off: smaller L versus larger L #

A smaller inductance reacts faster to load steps (good for fast AI-accelerator transients) but carries more ripple, which raises RMS loss and stresses the power stage. A larger inductance is calmer and more efficient but slows the loop and needs more board area. Most designs land the phase ripple somewhere around 20–40% of the phase current — enough margin for stable control without oversized coils.

Coupled inductors, where adjacent phases share a core, can cancel some ripple and shrink the solution. They are common on high-density accelerator cards, but they change the loop math and are not a drop-in swap for uncoupled parts. Treat them as a different component family, not a suffix variant.

Why AI-server and GPU core rails stress inductors differently #

AI accelerators push the inductor harder than almost any other load. The rail is often below 1 V, the total current runs into hundreds of amps across many phases, and the load steps in nanoseconds when a compute kernel starts. Three things follow for sourcing:

• Saturation margin must be checked at worst case, not nominal. A part that

is fine at steady load can clip during a transient and let the phase current overshoot the power stage's rating.

• Height is a hard constraint. Under a cold plate or between a mezzanine and

the board, the inductor's profile often decides whether a design fits at all.

• Thermal derating is real. DCR loss scales with current squared, so the

same part that is comfortable at 20 A can overheat at 40 A.

In the typical 48 V → 12 V → core hierarchy, the inductor that matters most for buyers sits on the final 12 V-to-core stage, paired with the smart power stage that does the switching. Representative switching devices in the LimChip catalogue include the Infineon TDA21590 and the Renesas ISL99390FRZ-TR5935 90 A smart power stages; the coil across each is the magnetic that this guide is about.

Sourcing and qualification checks #

Magnetics attract the same counterfeit and mislabelling risks as any passive, and the failure mode is subtle: an unmarked or relabelled coil may read the right inductance on a cheap LCR meter yet collapse early under saturation or run hot. Before committing stock:

• Request the vendor datasheet and confirm Isat and Irms at temperature, not

just the room-temperature headline.

• For automotive or ruggedised builds, check AEC-Q200 qualification where

applicable.

• Confirm package, height, terminal style and reel/tray packing against the

board and pick-and-place setup.

• Verify date code and lot consistency across a batch; mixing magnetics

grades changes loop behaviour phase to phase.

• Treat a second source as a re-qualification, not a copy. Different L and DCR

tolerances shift the control loop, so a "same inductance" substitute is not automatically drop-in.

Practical selection checklist #

1. Start from the phase current and Vout, not the inductance. 2. Pick L so ripple lands near 20–40% of phase current at the chosen fsw. 3. Confirm Isat clears the worst-case load step with margin. 4. Confirm Irms clears the steady thermal load at the ambient you will ship in. 5. Minimise DCR where efficiency or self-heating is tight, accepting more size. 6. Lock the height and shielding against the mechanical envelope. 7. Verify the part against a real datasheet and lot before releasing an RFQ.

Conclusion #

The output inductor is the component that turns a switching power stage into a usable core rail, and on AI-server and GPU boards it is loaded harder than almost anywhere else. Selecting one is less about hitting a microhenry number and more about reading saturation current, RMS current and DCR against the real load, temperature and height envelope. Get those three right, qualify the part against its datasheet and lot, and the rest of the VRM has room to do its job.