Skip to content

Dead time of the power stage

The two MOSFETs of a half-bridge must never conduct at the same time: with both on, the bus is shorted through the leg and the current is limited only by the resistance of the switches and the layout. This shoot-through destroys the MOSFETs within a few edges. The timer therefore delays every turn-on by a fixed dead time after the turn-off of the other switch of the same leg.

The user manual of the B-G431B-ESC1 (UM2516) does not give a dead time, and a value copied from another board is not safe: it depends on the gate circuit. This page derives the value from the schematic and a circuit simulation, step by step, so that anybody can repeat it, for this board or for another one.

Result

dead-time-ns = <1200> in app/boards/b_g431b_esc1.overlay. The simulation shows shoot-through up to 600 ns and a first clean value at 900 ns; 1200 ns adds a margin of 1.3. Confirm it on the board before the first motor run (bring-up step 1).

Where the dead time is set

Layer What happens
Devicetree dead-time-ns of the power stage node in app/boards/b_g431b_esc1.overlay. 0 means "not set": the driver then refuses to enable the outputs
Driver drivers/power_stage/power_stage_stm32.c converts it to timer clock periods, rounded up: 1200 ns × 170 MHz = 204 ticks (5.88 ns each)
Timer TIM1 inserts it on every complementary pair CHx / CHxN through the DTG field of TIM1_BDTR. 204 ticks need the second DTG range, (64 + DTG[5:0]) × 2 ticks, so DTG = 0x80 | 38 = 0xA6
Log At startup: TIM1: 170000000 Hz clock, ARR 4250 (20000 Hz), dead time 1200 ns (204 ticks)

The gate driver itself, an L6387ED, adds no dead time. Its interlocking only keeps both outputs low when both inputs are high, which does not protect against a slow turn-off.

What limits the dead time

The timer dead time starts at the command to switch off. The MOSFET is off later, when its gate has discharged below the threshold VGS(th). Two effects decide whether the time between is long enough:

  1. Turn-off time. The gate charge leaves through the sink stage of the driver and the gate resistor. The larger the gate resistor, the longer the turn-off, and the longer the dead time must be.
  2. Miller bump. When the other switch turns on, the phase node moves by the bus voltage in a few tens of nanoseconds. The gate-drain capacitance Cgd of the switch that is off couples this edge into its gate. The resulting current through the gate resistor lifts the gate voltage for a moment. If the gate has not fully discharged yet, the bump starts from a higher level and can cross VGS(th): a short shoot-through at every edge, although the gate signals never overlap.

So three criteria must hold, checked in this order:

Criterion Meaning
Gate gap > 0 The off gate is below VGS(th) before the other gate rises above it
Miller bump < VGS(th) The off gate stays below the threshold at the edge of the other switch
No extra bus current The peak current drawn from the bus is only the charge of the switch node (Coss, snubber), not a short through the leg

The first criterion is the one a datasheet calculation gives. On this board the second one is the stricter one, as the results below show.

Step 1: the gate circuit from the schematic

The schematic of the board, MB1419 rev C, is part of the B-G431B-ESC1 documentation on st.com (file mb1419-g431cbu6-c01_schematic.pdf). Sheet 5, POWER, contains everything that sets the switching times. The three legs are identical; leg U is shown.

Leg U of the B-G431B-ESC1 (MB1419 rev C, sheet 5): TIM1_CH1 and TIM1_CH1N drive the L6387ED U10, which drives the high-side MOSFET Q2 through R34 (33 Ω) and the low-side MOSFET Q3 through R38 (33 Ω). R35 and R39 (10 kΩ) hold the gates low. D6 and C43 (1 µF) form the bootstrap supply. R37 (1 Ω) and C51 (10 nF) are a snubber from the phase node to ground. Schematic excerpt © STMicroelectronics, Open Platform License Agreement

Leg U of the B-G431B-ESC1 (MB1419 rev C, sheet 5): TIM1_CH1 and TIM1_CH1N drive the L6387ED U10, which drives the high-side MOSFET Q2 through R34 (33 Ω) and the low-side MOSFET Q3 through R38 (33 Ω). R35 and R39 (10 kΩ) hold the gates low. D6 and C43 (1 µF) form the bootstrap supply. R37 (1 Ω) and C51 (10 nF) are a snubber from the phase node to ground. Schematic excerpt © STMicroelectronics, Open Platform License Agreement

Auxiliary supply (MB1419 rev C, sheet 5): the L7986 buck converter U9 makes the +10V rail that supplies the gate drivers. The feedback divider R30 (4.7 kΩ) and R33 (300 Ω) sets 0.6 V × (1 + 4700 / 300) = 10 V. Schematic excerpt © STMicroelectronics, Open Platform License Agreement

Auxiliary supply (MB1419 rev C, sheet 5): the L7986 buck converter U9 makes the +10V rail that supplies the gate drivers. The feedback divider R30 (4.7 kΩ) and R33 (300 Ω) sets 0.6 V × (1 + 4700 / 300) = 10 V. Schematic excerpt © STMicroelectronics, Open Platform License Agreement

Low-side shunts (MB1419 rev C, sheet 5): R54, R55 and R56, 3 mΩ each, between the source of each low-side MOSFET and ground. Schematic excerpt © STMicroelectronics, Open Platform License Agreement

Low-side shunts (MB1419 rev C, sheet 5): R54, R55 and R56, 3 mΩ each, between the source of each low-side MOSFET and ground. Schematic excerpt © STMicroelectronics, Open Platform License Agreement
Value Parts (legs U, V, W) Value on MB1419 rev C Effect on the dead time
Gate driver U10, U11, U13 L6387ED, interlocking, no dead time of its own Delay about 105 to 110 ns, sink about 650 mA
Driver supply +10V rail (U9, R30, R33) 10 V Gate voltage when on; less charge to remove than at 12 or 15 V
Gate resistor R34, R38 / R40, R43 / R47, R50 33 Ω, no turn-off diode Sets the turn-off time. The largest effect
Gate pull-down R35, R39 / R41, R45 / R48, R52 10 kΩ Keeps an unpowered gate off; negligible for the timing
MOSFET Q2 to Q7 STL180N6F7 Gate charge and Cgd (Miller bump)
Bootstrap D6, C43 / D8, C52 / D9, C58 STPS0560Z, 1 µF High-side supply; not timing relevant
Snubber R37, C51 / R44, C59 / R51, C65 1 Ω + 10 nF to ground Slows the edge of the phase node at low current
Shunt R54, R55, R56 3 mΩ Part of the leg; small effect
Input filter C47, C48 / C53, C54 / C60, C61 100 pF, no series resistor Negligible delay at the MCU pin

Step 2: a first estimate by hand

Below the Miller plateau, the gate of the switch that turns off discharges like an RC circuit, through the sink of the driver (about 18 Ω, from the 650 mA sink current) and the gate resistor, into the input capacitance Ciss (4825 pF at 25 V in the datasheet):

tau      = (R_sink + R_g) * C_iss = (18 + 33) Ohm * 4.825 nF = 246 ns
t(10 V -> 3 V) = tau * ln(10 / 3)                          = 296 ns
t_off   >= t_driver + t(10 V -> 3 V) = 105 ns + 296 ns      ≈ 400 ns

The same estimate with 10 Ω gives about 270 ns: the 33 Ω resistors make the turn-off about 1.8 times slower. That explains why a dead time chosen for a board with small gate resistors is not safe here.

This is a lower bound and not a dead time:

  • Ciss is larger than the datasheet value at 25 V, because the switch that turns off has almost no drain-source voltage at first, and Cgd rises steeply at low VDS.
  • The Miller plateau is missing: while the phase node moves, the gate stays at about 3.1 V and the gate current, about 3.1 V / 51 Ω = 61 mA, only charges Cgd. Its length is Qgd / I_gate and depends on the current and the snubber.
  • The Miller bump of the other edge cannot be estimated reliably by hand.

The waveform below shows these phases: the RC discharge from 110 to 600 ns, the plateau from 600 to 950 ns, and only then the drop below VGS(th). The real turn-off takes more than twice the hand estimate. A simulation is needed.

Step 3: the ngspice sweep

The ngspice model of one leg, sfs-stepper-power-spice (Simulated motor), contains:

Model Content
models/l6387_behav.lib L6387E behavioural: interlocking, about 107 ns delay, 30 Ω source, 18 Ω sink, bootstrap
models/stl180n6f7_sfs.lib STL180N6F7 as a VDMOS model fitted to the datasheet: Rds(on), body diode, Ciss / Coss / Crss, Qg
circuits/esc1_leg.lib One leg: driver, two MOSFETs, gate resistors, bootstrap, shunt, layout inductances

tools/power/deadtime_sweep.py uses this model with the values of step 1 and adds the parts the leg model does not contain yet: the 10 kΩ gate-source resistors and the snubber. For each timer dead time from 300 to 1500 ns in steps of 100 ns, at junction temperatures of 25 °C and 100 °C, it simulates the leg at 24 V, 20 kHz, duty 0.5, with phase currents of −3, −0.25, +0.25 and +3 A, and keeps the worst case of the three criteria.

# ngspice (42 or newer) in PATH, or NGSPICE=<path to ngspice_con.exe>
python tools/power/deadtime_sweep.py sweep        # about 15 minutes -> build/deadtime/sweep.csv
python tools/power/deadtime_sweep.py plot         # the figures of this page
python tools/power/deadtime_sweep.py schematic --pdf mb1419-g431cbu6-c01_schematic.pdf

In the west workspace the model is found at modules/lib/sfs-stepper-power-spice. Elsewhere, pass --spice <path> or set SFS_POWER_SPICE. The board values are the BOARD table at the top of the script, with the schematic reference of each value.

Step 4: results

Sweep of the timer dead time: left, the highest gate voltage of the switch that is off while the other switches; right, the peak current from the bus with a 3 A load, logarithmic. Below 900 ns the Miller bump crosses VGS(th) = 3 V, and up to 700 ns the bus current shows a short through the leg

Sweep of the timer dead time: left, the highest gate voltage of the switch that is off while the other switches; right, the peak current from the bus with a 3 A load, logarithmic. Below 900 ns the Miller bump crosses VGS(th) = 3 V, and up to 700 ns the bus current shows a short through the leg
Timer dead time Gate gap, 25 / 100 °C Miller bump, 25 / 100 °C Peak bus current, 25 / 100 °C Verdict
300 ns gates overlap 4.61 / 4.68 V 111 / 88 A Shoot-through
400 ns 21 / 21 ns 4.30 / 4.35 V 66 / 52 A Shoot-through
500 ns 28 / 32 ns 4.02 / 4.07 V 39 / 31 A Shoot-through
600 ns 114 / 122 ns 3.79 / 3.82 V 22 / 17 A Shoot-through
700 ns 205 / 200 ns 3.61 / 3.65 V 11.4 / 9.5 A Short pulses at every edge
800 ns 281 / 277 ns 3.52 / 2.98 V 9.0 / 8.5 A Bump at the threshold
900 ns 365 / 361 ns 2.79 / 2.73 V 9.0 / 8.5 A First clean value
1000 ns 450 / 446 ns 2.60 / 2.54 V 9.0 / 8.5 A Clean
1100 ns 539 / 537 ns 2.45 / 2.39 V 9.0 / 8.5 A Clean
1200 ns 632 / 630 ns 2.34 / 2.27 V 9.0 / 8.5 A Chosen
1300 ns 726 / 724 ns 2.25 / 2.19 V 9.0 / 8.5 A Clean
1500 ns 918 / 916 ns 2.13 / 2.06 V 8.9 / 8.4 A Clean

The 9 A floor of the bus current is the charge of the switch node: the output capacitance of the MOSFETs and the 10 nF of the snubber, charged to 24 V within a few tens of nanoseconds. It does not change with the dead time, so it is not a short.

One edge at −3 A, 25 °C. Left, 500 ns: the low-side gate (blue) is still on its Miller plateau when the high-side gate (orange) rises above VGS(th); both switches conduct and the bus current reaches 20 A. Right, 1200 ns: the low side is off at about 950 ns, the high-side gate rises only after 1300 ns, and the bus current carries only the load. The dotted lines are the command off of the low side and the end of the timer dead time

One edge at −3 A, 25 °C. Left, 500 ns: the low-side gate (blue) is still on its Miller plateau when the high-side gate (orange) rises above VGS(th); both switches conduct and the bus current reaches 20 A. Right, 1200 ns: the low side is off at about 950 ns, the high-side gate rises only after 1300 ns, and the bus current carries only the load. The dotted lines are the command off of the low side and the end of the timer dead time

Step 5: the chosen value and the margin

The simulation gives 900 ns as the smallest dead time that meets all three criteria at 25 °C and at 100 °C. The firmware uses 1200 ns, about 1.3 times that, because the model is not exact in the directions that make the real minimum larger:

  • VGS(th). The criterion uses 3 V. The threshold of a real MOSFET spreads between parts and falls when it gets hot. The fitted model hardly changes its threshold with temperature, which is why the 25 °C and 100 °C curves are so close. A lower threshold raises the minimum dead time.
  • Driver delay mismatch. The model gives both channels of the L6387E the same delay. A real driver turns one channel off later than it turns the other on, and the difference uses up part of the margin.
  • MOSFET model. The model is fitted to the datasheet by this project; it is not the model of the manufacturer.

The price of a longer dead time is small at 20 kHz. The phase voltage error grows by about 1.2 µs / 50 µs = 2.4 % of the bus voltage per leg, and the current loop corrects it. The largest leg duty at which the shunt current can still be sampled falls from about 0.95 to 0.93. The shared-leg drive of this motor works well below that.

Step 6: check on the board

A simulation narrows the choice. It does not replace a measurement. Before connecting a motor:

  1. Supply at 12 V with a current limit of about 200 mA. No motor connected.
  2. Torque Enable over the bus. Without a load the current loop cannot reach its set point, so a leg may sit at a duty of 0 or 1 and show no edges. In that case connect a power resistor (for example 10 Ω, 10 W) between U and W in place of the motor.
  3. Oscilloscope on the low-side VGS of one leg (gate and source of Q3, Q5 or Q7) and on the phase node. Use a differential probe for the high side, never a ground clip on the phase node.
  4. The off gate must stay well below VGS(th) at every edge, and the supply current must stay small and steady. A current limit that trips, or MOSFETs that get warm, mean the dead time is too short: switch off at once.

Only after this measurement, and not below 900 ns, can the dead time be lowered. The full procedure is bring-up step 1, and hardware checklist check 5.

With a motor, the firmware can also measure the effective dead time: the voltage each leg loses during the dead time, from the phase voltage the current loop needs at a given current (Measurements with MCUViewer). It shows how large the loss really is and how it distorts the phase voltage. It does not replace the oscilloscope above, because it cannot see the gates or a short shoot-through.

Limits of the method

  • The sweep covers one operating point: 24 V, 20 kHz, phase currents up to ±3 A. Repeat it for a higher bus voltage or a different gate circuit.
  • Layout inductances (3 nH commutation loop, 1 nH shunt) are assumptions of the model, not values from the board layout.
  • The leg voltage error table that the simulated motor uses (Simulated motor) is still computed for 500 ns and 10 Ω. It is updated when the leg model of sfs-stepper-power-spice takes over the values of step 1.
  • The dead time compensation of Stepper-C is not implemented yet (dead_time_ns is reserved). When it is, it must use the effective dead time seen in the phase voltage, not the 1200 ns of the timer.

For another board

  1. Read the gate circuit from its schematic: driver and its supply, gate resistors, turn-off diodes, MOSFET, snubber, shunt (step 1).
  2. Make the hand estimate of step 2 to see the order of magnitude.
  3. Enter the values in BOARD of tools/power/deadtime_sweep.py. For another driver or MOSFET, add its model to sfs-stepper-power-spice first.
  4. Run the sweep, find the first value that meets all three criteria, add a margin, and measure it on the board before connecting a motor.