PSFB Lagging-Leg Dead Time: Why a Longer Delay Loses ZVS
By the Power Electronics Hub editor · Published 2026-11-07
A 1.5 kW phase-shifted full bridge takes a 400 V bus down to 48 V through a 5:1 transformer at 100 kHz. The leakage plus shim inductance is 15 µH and each MOSFET has about 200 pF of energy-equivalent output capacitance, so the lagging-leg node carries 400 pF. By the usual energy check the lagging leg keeps zero-voltage switching down to 31 % load. With the lagging dead time set to 200 ns it actually stops at 76 %, and with 300 ns it never gets ZVS at all, not even at full load. Nothing about the energy changed. The dead time was simply too long.
The window has a closing edge
When the lagging-leg switch turns off, the transformer is shorted by the conducting rectifiers and only L_r keeps the primary current going. That current swings the switch node through C_eq = 2·C_oss + C_tr as a quarter of an L_r–C_eq resonance. If the energy is sufficient the node reaches the opposite rail at t1 and the incoming switch's body diode takes over. From then on the full bus voltage sits across L_r with the polarity that drives its current down, so the diode current falls linearly and reaches zero at t2.
What happens after t2 is the part that is easy to forget. The current reverses, the diode stops conducting, and the node starts resonating back toward where it came from. Turning the gate on any time after t2 means turning on into a voltage that is rising again. A dead time is therefore not a margin that can only help when it is longer. It has to land between t1 and t2.
- Energy condition
- ½·L_r·I_lag² ≥ ½·C_eq·V_in² ⇔ I_lag ≥ V_in / Z0, Z0 = √(L_r / C_eq)
- Node reaches the rail
- t1 = asin(V_in / (I_lag·Z0)) / ω_r, ω_r = 1/√(L_r·C_eq)
- Diode current reaches zero
- t2 = t1 + L_r·I_lag·cos(ω_r·t1) / V_in
- Voltage at turn-on after t2
- V_ds = V_in·(1 − cos(ω_r·(t_d − t2))), until the other diode clamps at V_in
- Turn-on loss of the leg
- P ≈ ½·C_eq·V_ds² × 2 edges × f_s = C_eq·V_ds²·f_s
In the example the lagging current at full load is 6.37 A: the output-inductor valley of 28.85 A divided by 5, plus 0.6 A of magnetizing current. The node reaches the rail at 25.6 ns and the diode current is gone at 251.5 ns. As the load drops, t1 moves later and t2 earlier, and at the energy limit, 9.73 A of output current, they meet at the quarter period, (π/2)·√(L_r·C_eq) = 122 ns.
What a fixed dead time costs across load
The chart takes the same converter and plots the lagging-leg turn-on loss, C_eq·V_ds²·f_s, against load for three fixed dead times: the quarter period and two longer values. The worst case for this leg, a full hard turn-on from 400 V, is 32 µJ per edge or 6.4 W.
With 200 ns the loss starts at 76 % load. At half load the switch turns on with 89 V across it and the leg loses 0.32 W. At 30 % it is 195 V and 1.5 W, and at 10 % load, where the converter delivers 150 W, it is 323 V and 4.2 W, close to 3 % of the output. With 300 ns the diode current has already reversed at full load, the switch sees 76 V, and below about 65 % load the node has swung all the way back, so every lagging turn-on is a full 400 V hard turn-on.
With the quarter period, 122 ns, the loss stays at zero all the way down to the 31 % energy limit, and below that it rises more slowly than either fixed alternative: 0.7 W at 20 % load and 2.6 W at 10 %.
Below the energy limit the quarter period is still the best choice
Once the energy runs out the node no longer reaches the rail. It peaks at I_lag·Z0 after exactly a quarter period and then swings back. Turning on at that moment is valley switching: the residual voltage V_in − I_lag·Z0 is the lowest the switch can see, and any other dead time adds to it. So in this lossless picture one fixed dead time, the quarter period, is right at every load. It gives ZVS wherever the energy allows it and the smallest turn-on voltage everywhere else.
This holds for a linear capacitance. Superjunction MOSFETs have a C_oss that rises steeply at low drain voltage, so the real transition is not a sinusoid and the 'quarter period' computed from an energy-equivalent C_oss is only a starting point. The time-equivalent C_o(tr) from the datasheet is the better input for timing.
What a shorter dead time costs instead
The price of switching at the quarter period at full load is body-diode conduction: the diode carries current from t1 = 26 ns to 122 ns. With an illustrative forward voltage of 0.9 V that is 0.08 W for the lagging leg at full load and almost nothing at 30 %. Superjunction body diodes also have reverse-recovery charge, which this estimate leaves out, so a controller that shortens the lagging delay at heavy load has a reason to do so. Still, the amounts are a tenth of a watt against the watts lost by turning on after t2.
The leading leg is the opposite case. Its transition is driven by the reflected load current, which does not reverse during the dead time, so there is no closing edge, only a minimum: 22 ns at full load, 38 ns at half load, 94 ns at 10 % and 115 ns at 5 %. Its dead time has to grow toward light load, while the lagging leg's should stay near the quarter period. A single dead time for both legs fits only by luck. Controllers such as TI's UCC28950 keep the two legs' delays separate, and its datasheet says the delay may be adjusted dynamically according to the load level through the ADEL pin.
Finding the quarter period on the bench
- 1Start with the quarter period from L_r and the time-equivalent C_oss plus transformer capacitance, using the PSFB ZVS calculator on this site.
- 2Measure the lagging-leg drain-source voltage at a light load where ZVS is marginal, and step the lagging dead time in small increments.
- 3Look for the dead time where the voltage at turn-on is lowest. Too short and the gate rises on a falling edge; too long and the node has already started back up.
- 4Repeat at full load to check that the same value still lands after t1, and set the leading-leg delay separately at the lightest load that matters.
Probing matters here for the same reason it does in any dead-time measurement: the node moves hundreds of volts in tens of nanoseconds, so a long ground lead will paint ringing on the trace that is not really there.
Sources
- Texas Instruments, SLUA560, "Designing a UCC28950 600-W phase-shifted full-bridge converter" (2·C_OSS at the switch node, shim inductor sized for ZVS down to 50 % load).
- Texas Instruments, UCC28950/UCC28951 datasheet, SLUSDB2B (separate delays for the two legs, adjustable with load through ADEL).
- J. A. Sabaté, V. Vlatković, R. B. Ridley, F. C. Lee and B. H. Cho, "Design considerations for high-voltage high-power full-bridge zero-voltage-switched PWM converter," IEEE APEC 1990.
- All numbers in this article were computed in Python from the equations above and checked against a numerical integration of the L_r–C_eq transition. Body-diode forward voltage is an illustrative value; reverse recovery and C_oss nonlinearity are not included.
This guide explains general engineering principles. Always verify your design with datasheets, simulation and measurement.