Server Power Supply Rectifier Design: Thermal-First Thinking at 40 A

Server PSU rectifier design at 40 A: the loss arithmetic, why the package and copper set the rating, and what hot-swap demands of the output stage.

Request a Quote

Server Power Supply Rectifier Design: Thermal-First Thinking at 40 A

Server power supplies run at high power density, and the output rectifier is where the thermal design is won: 40 A of load current turns every 0.1 V of forward drop into 4 W of heat inside a rack full of sealed boxes. This guide walks the thermal-first method, the package decision, the loss math, and the redundancy and surge checks.

Last updated: October 11, 2026

Power Density Forces Thermal-First Design

Server platforms push power density higher with every generation—more watts per liter, more watts per square centimeter of board—and the output rectifier sits at the heart of both limits. The design order follows the heat: define the loss budget first, then the package and heat path, then the part. A rectifier that passes the electrical check but cannot shed its watts fails the density test before it fails electrically.

The thermal-first habit also changes the datasheet reading: the number that matters is the loss at the hot condition and the case temperature the assembly can hold, not the 25 °C headline current.

The density trend also explains why the output stage is where the wins are found: the input and PFC stages were optimized in earlier generations, and the output rectifier’s fixed drop at high current is one of the last large, addressable loss terms. A 40 A output turns every 0.1 V of drop into 4 W—a lever the rest of the supply cannot match.

The same density pressure appears at the system level: a rack of servers moves the heat to a facility cooling system, and every watt saved in the supply is a watt the facility does not have to remove. The supply’s efficiency is the data center’s thermal budget.

High-Current Output Rectification: D2PAK and TO-220 Territory

At 20–40 A output, two packages dominate: D2PAK for board-level mounting, where the copper and vias are the heatsink, and TO-220 for bolt-on mounting, where a real heatsink carries the heat. The choice follows the mechanical architecture—a sealed module with no heatsink access favors D2PAK, a chassis with airflow and a heatsink favors TO-220—and the comparison is covered in the package decision article. For the 40 A output of a server supply, a common architecture splits the current across parallel rectifier paths, each carrying 20 A or less, so the thermal path per device stays manageable.

The parallel split is not just a thermal convenience—it also lowers the per-device conduction loss and spreads the surge duty, so the redundant architecture and the thermal architecture are the same decision. The package choice and the split count are made together, before the part number is selected.

Calculating Losses at 40 A Output

Worked check for a 40 A output split across two 20 A rectifier paths. Each device carries 20 A; using the AMBR40250S with its 0.90 V maximum forward drop at 20 A and 25 °C, the first-pass loss per device is about 18 W, 36 W total. The first pass is the starting estimate only: the final budget uses the maximum VF at the operating junction temperature, adds the leakage term at the working reverse voltage, walks the full thermal chain—interface, heatsink or copper, ambient—and is confirmed by a case-temperature measurement on the prototype, because VF and leakage both move with temperature.

Load condition	First-pass loss (typ/max at 25 °C)	Final budget rule
20 A per device	0.83 typ / 0.90 max → ~16.6–18 W	Max VF at hot TJ + leakage
40 A total (2 × 20 A)	~33–36 W	Full thermal chain + prototype

The table shows the method, not a finished number: the 25 °C typical figures are where the conversation starts, and the measurement is where it ends. The same method scales to a 60 A output by adding a third path, keeping the per-device current and temperature constant while the total rises.

The leakage term belongs in the same final budget: at the working reverse voltage and the hot junction, the 10 mA maximum leakage figure of the AMBR40250S class can add real watts when the node is continuously biased, and it must share the same heat path as the forward loss. The net budget—forward plus leakage, iterated to the hot condition—is the number the heatsink is sized for.

The thermal chain is the same for both packages: junction-to-case from the datasheet, interface from the mounting, and heatsink-to-ambient from the enclosure. The D2PAK’s interface is the solder and copper; the TO-220’s is the TIM and torque—the difference is which chain the design can control.

The Loss Arithmetic at 40 A

At 40 A the output rectification loss decides the whole thermal design, and it can be estimated from published figures before anything is placed. A diode at a 0.8 V average forward drop dissipates 32 W at 40 A; a power MOSFET at 1.4 mΩ dissipates 2.24 W. The same comparison at 12 A is 9.6 W against 0.2 W, so the argument holds across the range rather than at one operating point.

Decision	Figure behind it	Consequence
Diode or synchronous switch	Forward drop against on-resistance at the output current	Tens of watts against a couple of watts
Package	Junction-to-case and junction-to-ambient resistance	The same die is rated 5× apart on the two references
Copper and airflow	Ambient-referenced rating at a stated pad area	Sets whether the ambient figure describes the real board
Surge path	Non-repetitive surge and avalanche ratings	Hot-swap and inrush events, not steady-state current

The package row is where 40 A designs usually fail first. onsemi’s NTMFS5C430NL is rated 200 A continuous at a 25 °C case but 38 A at a 25 °C ambient on a 650 mm² two-ounce copper pad, with 1.4 °C/W junction-to-case against 40 °C/W junction-to-ambient. A 40 A output stage therefore sits right at the boundary of the ambient-referenced rating of a single device in that package — which is exactly why server supplies use several devices in parallel, spread the loss, and put the copper underneath them.

Two boundary conditions belong in the output-stage design. Paralleled rectifiers share current only when they are matched and thermally coupled, so the sharing depends on the datasheet tolerance rather than on an assumption of equality. And hot-swap adds a surge event that the steady-state loss calculation does not describe at all: the non-repetitive surge and avalanche ratings govern that path, and they are separate numbers from the continuous current. The sharing mechanics are covered in Paralleling Rectifier Diodes, and the package comparison in Diode Rectifier vs Synchronous Rectification.

Redundancy, Hot-Swap, and Surge Considerations

Server supplies add two duties the output rectifier must survive. Redundancy (N+1) means a failed unit is replaced without powering down, and the hot-swap event exposes the input and output stages to the inrush of recharging bulk capacitors—the surge column must be checked against that real waveform, not quoted as a headline. The redundant architecture also shares the load across units, so the rectifier’s derating at the hot condition is part of the availability math.

The surge check follows the five-step method in the surge rating guide: define the event, read the rating at the matching width and starting temperature, add margin for repeated hot-plug cycles, and confirm the thermal budget at the same conditions.

The N+1 sharing also changes the thermal derating: a unit carrying 110% of its share during a peer failure runs the rectifier hotter, and the design margin must cover the degraded state, not just the balanced one. The availability analysis and the thermal analysis read the same derating curve.

The hot-swap event is also a test moment: the replacement unit powers on into a live bus, and the inrush of the empty bulk capacitors is the surge the design must survive on every service event, not once at the factory. The measurement protocol for the 40 A output is fixed—same load, same ambient, same soak, and the case temperatures of all parallel devices recorded together—so the sharing and the margin are read from one table.

Layout Notes for Parallel Output Rectifiers

Parallel rectifier paths work only when they share current evenly, and layout is half the sharing. Symmetric traces from the transformer output to each rectifier keep the parasitic resistance equal; a common copper area and thermal vias keep the temperature equal; and a small balancing resistance or matched part selection handles the residual difference. The parallel design rules live in the paralleling guide; the server takeaway is that the heat path and the electrical path are designed together.

The measurement closes the loop: case temperatures within a few degrees across the parallel paths confirm the sharing, and a large spread points back to the layout or the part matching. The thermal camera and the thermocouple are the sharing instruments.

Engineering note. The 40 A example uses the AMBR40250S’s datasheet forward drop at 20 A and 25 °C as the first-pass estimate; the final loss budget must use the maximum VF at the operating junction temperature, add leakage at the working reverse voltage, and be verified by measurement, because a single-pass 25 °C calculation is an estimate, not a thermal design. The two-path split is a common server architecture, not a fixed requirement.

FAQs

Why is thermal design first in a server PSU?

Because power density makes heat the binding limit. A rectifier that passes electrically but cannot shed its watts fails the density test before it fails electrically.

How much loss does 40 A of output current create?

At 20 A per device with a 0.90 V maximum drop, about 18 W per device, 36 W total as a first pass. The final budget uses the hot-condition values and the full thermal chain.

D2PAK or TO-220 for a server output?

The mechanical architecture decides: sealed modules with board-level cooling favor D2PAK, chassis designs with a heatsink and airflow favor TO-220. The package comparison article draws the boundary.

What does hot-swap do to the rectifier?

It exposes the stage to the inrush of recharging bulk capacitors; the surge column must be checked against that real waveform with margin for repeated events.

How do I make parallel rectifiers share current?

Symmetric layout, equal trace resistance, a common thermal area, and matched parts. The paralleling guide covers the full method.

Conclusion

Server output rectification is thermal-first: set the loss budget, choose the package for the mechanical architecture, split the current across manageable paths, and verify with a hot-condition measurement. The surge and hot-swap checks close the availability story.

Review the AMBR40250S product page on the Good-Ark site, and contact Good-Ark with your output current, mechanical layout, and thermal conditions for a rectifier recommendation.