The Drive Stage
The Drive Stage: Designing a BLDC Motor Driver That Does Not Die
A BLDC driver is five sections, six switches, and one rule that matters more than the rest: the two devices in a half-bridge must never conduct together. This is how the drive stage works, how it dies, and the arithmetic that keeps it alive, with a worked 48 V design.
The microcontroller is the brain. It decides which phase gets current, and how much. But a brain cannot push a door open. The gate driver is the hands, it takes the microcontroller's 3.3 V whisper and turns it into the 12 V shove that opens a MOSFET. And the MOSFET is the door: when it opens, current flows from the battery into the motor winding, and the motor turns.
There are six doors, in three pairs. Each pair is a corridor with a door at the top and a door at the bottom. The rule that matters more than any other in this article is simple: the two doors in one corridor must never be open at the same time.
If they are, the battery is shorted straight to ground through two pieces of metal that are barely more resistive than a wire. The current is limited only by the inductance of the loop, and the MOSFETs die in microseconds, before any fuse or software can react.
That failure has a name, shoot-through, and it is the first of five ways a motor driver dies. The other four are quieter: a voltage spike that exceeds the MOSFET's rating, a bootstrap capacitor that runs dry at full duty, heat that feeds on itself as the on-resistance climbs, and a gate that turns itself back on through the Miller effect. Every one of them is a design decision, made on paper, weeks before the board is built. This article is the arithmetic for those decisions.
The words, in plain English
Three-phase inverter. The circuit that turns a DC supply into the three-phase current a brushless motor needs. Six switches, in three pairs, one pair per phase, U, V, and W. Each pair connects its phase either to the positive rail or to ground, and the controller switches them in sequence to make a rotating magnetic field.
Half-bridge. One pair of switches, one high-side, one low-side, driving one phase. The high-side switch connects the phase to the positive rail; the low-side connects it to ground. The point where the two meet is the switch node, and it swings from ground to the full rail every switching cycle.
High-side / low-side. The high-side MOSFET sits between the positive rail and the phase; the low-side sits between the phase and ground. The low-side is easy to drive, its source is at ground. The high-side is the hard one, its source rides up and down with the phase voltage, so its gate has to ride up and down with it, on a supply that floats.
Shoot-through. Both switches in one half-bridge conducting at the same time. The DC bus is shorted to ground through two on-resistances. The current is enormous, thousands of amps in theory, and the devices fail before anything else can intervene. It is the failure dead time exists to prevent.
Dead time. A short delay, tens to hundreds of nanoseconds, inserted between turning one switch off and turning the other on, so the two are never on together. It is a pause with both doors shut. Too little and the bridge shoots through; too much and the output voltage is wrong, and the motor vibrates at low speed.
Gate charge (Qg). The amount of charge the driver must push into a MOSFET's gate to turn it on, measured in nanocoulombs. The gate is a capacitor; switching it means charging and discharging that capacitor, every cycle, at the switching frequency. The gate drive power is Qg × Vdrive × frequency.
Miller plateau. The flat part of the gate voltage waveform during switching, where the gate voltage stops rising while the drain voltage swings. The charge that moves during the plateau, Qgd, the gate-drain or Miller charge, is what sets the switching time, not the total gate charge. During the plateau the driver pushes current at (Vdrive − Vplateau) ÷ Rtotal on turn-on, and at Vplateau ÷ Rtotal on turn-off, which is why turn-off is slower than turn-on for the same resistor.
Bootstrap. The cheap way to power a high-side gate driver: one diode and one capacitor. While the low-side switch is on, the switch node sits at ground and the supply charges the bootstrap capacitor through the diode. When the high-side turns on, the node flies up to the rail, the diode blocks, and the capacitor floats up with the node, supplying the high-side gate. The catch: it only recharges while the low-side is on, so a bootstrap driver cannot run at 100% duty cycle.
Charge pump. An on-chip voltage multiplier that generates the floating high-side supply without needing the low-side to conduct. It is what lets a driver like the DRV8323 or the DRV8305 run at 100% PWM duty cycle. It costs more silicon than a bootstrap diode; it buys back the top of the duty range.
DC link. The DC bus between the power supply and the inverter, and the capacitor across it. The DC-link capacitor's job is to absorb the switching ripple current the inverter draws in pulses, and to hold the bus voltage steady during load steps. It is sized by ripple current first and capacitance second, because a capacitor that overheats from ripple current fails long before its capacitance matters.
Rds(on). The drain-to-source on-resistance of a MOSFET, the resistance of the channel when the device is fully on. Conduction loss is I² × Rds(on), so it scales with the square of the current. And it rises with temperature, typically 40–80% higher at 125–150 °C than at 25 °C, which is why heat feeds on itself.
Switching loss. The energy lost every time a switch turns on or off, while voltage and current overlap. Roughly ½ × V × I × (t_on + t_off) × f_sw per device. It scales with frequency, which is why faster switching costs efficiency, and why the gate resistor is a trade-off, not a free choice.
Shunt. A small, precise resistor placed in the current path so the current can be measured as a voltage across it, V = I × R. A 1 mΩ shunt at 10 A gives 10 mV, which a current-sense amplifier multiplies up to something the microcontroller's ADC can read. The shunt's own loss is I² × R, so it is kept as small as the amplifier allows.
Current-sense amplifier (CSA). The amplifier that reads the tiny voltage across a shunt and scales it up. The DRV8323 and DRV8305 integrate three of them, with adjustable gain, 5, 10, 20, or 40 V/V on the DRV8323.
FOC (field-oriented control). The control method from Article #20: the controller transforms the three phase currents into a rotating frame and regulates torque and flux separately, with two PI loops. It needs accurate, continuous phase-current feedback, which is why where the shunts sit matters as much as how big they are.
PWM / duty cycle. Pulse-width modulation, the switch is on for a fraction of each cycle, and that fraction is the duty cycle. The motor sees the average. A 50% duty cycle at 48 V gives the motor 24 V on average.
VDS. The drain-to-source voltage across a MOSFET. The number that must never exceed the device's rating, not the bus voltage, but the bus voltage plus the overshoot from the switching loop's inductance.
UVLO (undervoltage lockout). The driver's self-protection: if its supply voltage falls below a threshold, it turns all the gates off rather than drive them weakly. A weakly driven MOSFET is a resistor, and a resistor in a power stage is a heater.
Commutation loop. The high-frequency current path through the DC-link capacitor, the high-side switch, the motor phase, the low-side switch, and back to the capacitor. It is the loop whose inductance decides the voltage overshoot at turn-off, and the loop whose area must be made as small as the layout allows.
Stray inductance. The inductance of every conductor in the circuit, traces, leads, bond wires, that is not a component on the schematic. It cannot be eliminated, only minimised. At turn-off it produces V = L × di/dt, and that voltage adds to the bus voltage across the switch that just turned off.
Ringing. The oscillation that follows a fast switching edge, when the loop inductance and the device capacitances form a resonant circuit. It shows up as voltage spikes, as radiated emissions, the subject of Article #19, and, at its worst, as a gate voltage that bounces high enough to turn a device back on.
The five sections
A BLDC driver has five functional sections, and every driver ever built is some version of this list:
- DC power supply and DC-link capacitor. The supply, the bulk capacitance that holds the bus up, and the ceramic capacitors placed close to the inverter that absorb the fast switching ripple.
- Three-phase MOSFET inverter. Six MOSFETs in three half-bridges, U, V, W, one high-side and one low-side per phase.
- Gate-driver circuit. The translator between the microcontroller's logic and the MOSFET gates: it provides the voltage, the current, the level shifting for the high side, and the dead time.
- Microcontroller or motor-control IC. The brain: it generates the PWM, runs the commutation and the FOC loops from Article #20, and handles faults.
- Position, current, voltage, and temperature feedback. Hall sensors or an encoder for position, shunts and amplifiers for current, a divider for bus voltage, and a temperature sensor on the power stage. Current feedback is used for overload detection, torque control, current limiting, and fault protection.
The controller determines the required switching state. The MOSFET bridge supplies the current to the motor windings. The gate driver is the part that stands between them, and it is the part most often designed as an afterthought, which is why it is the part that most often kills the board.
Why the microcontroller cannot drive the gates itself
The microcontroller normally should not drive the MOSFET gates directly in a serious power stage. A gate-driver IC provides the required voltage, current capability, level shifting, and protection functions between the controller and the MOSFETs.
The numbers make the reason plain. A microcontroller GPIO pin can source perhaps 20 mA. A power MOSFET with a total gate charge of 30 nC, switched in 16 ns, needs a peak gate current of about 1 A, fifty times what the pin can give.
With the pin alone, the switch would take microseconds to turn on instead of nanoseconds, and during those microseconds the MOSFET sits half-on, with full voltage across it and full current through it, dissipating power as heat. The switching loss would be fifty times higher. The device would cook at a switching frequency where a proper driver barely warms it.
The driver also solves the voltage problem. The microcontroller runs at 3.3 V. A power MOSFET's on-resistance is only guaranteed at a gate voltage of 10 V, drive it at 3.3 V and it is barely on, acting as a resistor. The driver translates 3.3 V logic to 10–15 V of gate drive, and it does it for the high side too, where the gate has to sit 10–15 V above a source that swings between ground and the full rail.
The two real gate drivers
Two Texas Instruments parts cover most of the ground this article's buyer will stand on, and their datasheets are the source for the numbers below.
DRV8323, a 6 to 60 V three-phase smart gate driver. It drives three high-side and three low-side N-channel MOSFETs, with adjustable slew-rate control, peak source current of 10 mA to 1 A and peak sink of 20 mA to 2 A, and it supports 100% PWM duty cycle through an integrated charge pump for the high side. Its adjustable dead times are 50, 100, 200, and 400 ns, set over SPI (the hardware version fixes 100 ns). It integrates three bidirectional current-sense amplifiers for low-side shunts, with adjustable gain of 5, 10, 20, or 40 V/V. It accepts 1.8 V, 3.3 V, and 5 V logic inputs, and it reports faults on a single nFAULT pin with detail in the SPI registers.
DRV8305, a 4.4 to 45 V three-phase smart gate driver, automotive-qualified (AEC-Q100). Three high-accuracy half-bridge drivers, 1.25 A and 1 A peak gate drive currents, programmable high- and low-side slew-rate control, and a charge-pump gate driver for 100% duty cycle. It integrates three bidirectional current-shunt amplifiers for accurate low-side current measurement, with variable gain and an adjustable offset reference, plus a 50 mA LDO (3.3 V or 5 V) to power the microcontroller. It runs 3-PWM or 6-PWM input control up to 200 kHz, uses automatic handshaking when switching to prevent shoot-through, and senses the VDS of both the high-side and low-side MOSFETs to protect them from overcurrent.
The older DRV8302 and DRV8300 use bootstrap diodes for the high side rather than a charge pump, which is the distinction the corrections section records. A charge pump costs more silicon; a bootstrap diode costs a duty-cycle limit. The choice is between the two, and the buyer should know which one their part uses before they ask it for 100% duty.
Dead time
Dead time is the delay between one switch turning off and the other turning on in the same half-bridge. Without it, the two devices overlap for the few nanoseconds it takes a MOSFET to turn off, and during that overlap the bus is shorted. With it, both devices are off for a short interval, and the load current freewheels through a body diode instead.
The DRV8323's settings, 50, 100, 200, 400 ns, are the right order of magnitude for a 20 kHz drive, where one switching period is 50 µs. Two hundred nanoseconds is 0.4% of the period. It sounds negligible. It is not, because the error it introduces does not shrink when the commanded voltage shrinks.
During dead time, both switches of a leg are off, and the output is decided by whichever diode carries the load current. If the current flows out of the leg, the lower diode conducts and the output sits at the negative rail, even though the controller wanted the upper switch on. If the current flows into the leg, the upper diode conducts and the output sits at the positive rail.
Either way the leg produces the voltage the current pushes it to, not the one that was commanded. Averaged over the period, the leg voltage is short by a fixed amount whose sign is set by the current direction:
Δv = −sign(i) × t_dead × f_s × V_dc
The worked example from the guides makes the scale concrete: a 400 V-class IGBT drive at 540 V DC link, 8 kHz carrier, 3 µs dead time loses about 13 V of average output voltage per leg, and the sign of that loss flips every time the phase current changes direction. At 50 Hz that is a few percent of the commanded voltage and mostly harmless. At 5 Hz, where the commanded voltage is a tenth as large, the same 13 V is a large fraction of it, and the current waveform flattens at every zero crossing.
The error is a square wave in phase with the current, not with the voltage: it contains every odd harmonic, and in a three-phase motor the 5th and 7th harmonic currents combine into torque ripple at six times the electrical frequency, vibration and audible noise at low speed, exactly where a robot arm does its precise work.
The fix is compensation, adding the lost voltage back into the reference before the PWM stage, or characterising the inverter at standstill and storing the error in a look-up table. But the fix needs the current's sign, and the current's sign needs accurate current sensing, which is why the sensing section below is not optional. And the dead-time error is one more reason the DRV8323's 50 ns setting exists: at 48 V and 20 kHz, the error at 200 ns is 0.19 V per leg, small, but at a 5 V commanded phase voltage it is nearly 4% of the command, and at 400 ns it is nearly 8%.
The dead time should be as short as the MOSFETs' turn-off allows, not as long as the datasheet's largest setting.
Choosing the switch: the voltage–frequency grid
MOSFETs and IGBTs have traditionally been used as the switches in inverter designs. MOSFETs can potentially switch at frequencies up to 100 kHz, but are more often used at speeds in the tens of kHz. As the voltage increases, the junction temperature rises and the reverse recovery of their internal diode deteriorates, causing increased heat and switching losses. At those higher voltage and current levels, IGBTs are usually the preferred choice, they operate at lower frequencies of around 20 kHz, and typical motor-control PWM frequencies for IGBTs are 10–20 kHz.
The grid, as the guides draw it:
Example Labs is a made-up company we use for worked examples.
- Below ~200 V, high frequency, silicon MOSFET or GaN. Low Rds(on), fast switching, low gate charge. This is where the Example Labs bench arm's 48 V drive lives, and where almost every robot joint drive lives.
- 200–600 V, moderate frequency, silicon carbide MOSFET against superjunction silicon MOSFET. SiC wins on switching loss; silicon wins on unit cost.
- 600–1,200 V, low frequency, IGBT, or SiC at the premium end. Conduction at high current, ruggedness, cost.
- 1,200–6,500 V, IGBT, with SiC entering. Traction, grid.
And the wide-bandgap step beyond it: silicon carbide handles high voltage and high temperature with low switching losses, which is why it has become the default for EV traction inverters; gallium nitride has the highest electron mobility of the three, switching well into the megahertz range with no reverse-recovery losses, and is strongest below roughly 650 V. A GaN or SiC device switches an order of magnitude faster than a silicon device at the same voltage, which is a gift to efficiency and a threat to the layout, because every nanosecond faster is more di/dt across the same loop inductance, and more overshoot.
A layout that performed acceptably with an older IGBT may produce excessive overshoot or ringing after a faster device is substituted.
The loss physics is the reason the grid is shaped the way it is. A MOSFET conducts through a resistive channel: conduction loss is I² × Rds(on), proportional to the square of the current. An IGBT conducts through a junction with a saturation voltage: conduction loss is roughly I × VCE(sat), proportional to current.
At low currents and voltages, MOSFET resistance is cheap and the IGBT's fixed drop is wasteful. At high currents, the IGBT's linear loss grows more gently than the MOSFET's squared loss, one reason IGBTs dominate above roughly 600–1,200 V and hundreds of amps. And the IGBT carries a tail current: stored charge in the drift region keeps current flowing after the gate turns off, making every cycle pay a fixed turn-off energy.
Practical ceilings sit around 20–50 kHz for standard IGBTs, hard-brick territory for a 100 kHz design.
Current sensing
Phase currents can be as high as 100 A in three-phase inverters operating from 12 to 60 V DC. To get the motor phase currents, shunts are typically placed at the DC-link return to ground, between the bottom switch and ground, or in-line with the three-phase power to the motor. Each placement has advantages and challenges, and the choice decides what the FOC loop can see.
In-line shunts sit in series with each phase, between the switch node and the motor. They give direct, continuous phase-current sensing, the current is visible in both switch states, and it can be averaged over one PWM cycle for higher accuracy, or the current controller can run twice per PWM. They can detect phase-to-phase and phase-to-ground shorts, and they can detect shoot-through. The price is that the amplifier needs a high common-mode input voltage with a high common-mode rejection ratio, because the shunt sits on a node swinging between ground and the full rail at the switching frequency. Accuracy is high.
Low-side shunts sit between the low-side MOSFET and ground, one per phase. The amplifier sees near-zero common mode, so a non-isolated current-sense amplifier, or the microcontroller's own integrated amplifier, is enough, and the system cost is lower. The price is that the sense is indirect and discontinuous: the phase current can only be measured when the low-side switch is on. It cannot be averaged over one PWM period. And at high duty cycle the low-side on-time shrinks toward zero, at 99% duty on a 20 kHz drive, the window is 0.5 µs, and the time the measurement needs (dead time, plus blanking, plus settling, plus ADC acquisition) is about 1.7 µs. The window closes before the measurement can finish. Accuracy is medium.
Single DC-link shunt sits in the return of the whole bus. It is the cheapest option and it measures the total current, but reconstructing three phase currents from one shunt over a PWM cycle is a puzzle that needs the switching state at every sample, and it cannot see a phase-to-phase short at all.
The practical rule: choose low-side shunts when the supplier can demonstrate valid sampling windows throughout the required speed, load, and modulation range. Consider in-line sensing when the low-side sampling windows restrict operation. And whichever placement is chosen, the fault-detection path, the VDS sensing and the gate disable, is qualified separately from the FOC feedback channel, because neither the topology nor the sensor isolation establishes a safe-torque-off capability on its own.
The DC link: the capacitor that takes the punches
The DC-link capacitor's job is to absorb the switching ripple current the inverter draws in pulses, and to hold the bus voltage steady during load steps. The selection has four layers: voltage rating, capacitance and ripple current, thermal management, and lifetime.
The voltage rating is chosen for the worst-case transient across the bus, not the average, because the stress on the film is proportional to the peak voltage it sees. Industry practice is to rate the capacitor at least 10–20% above the steady-state bus level: a 900 V bus means an 1,100 V or 1,200 V part. For a 48 V drive, a 63 V or 80 V part is the equivalent margin, and the margin matters more than it looks, because the overshoot from the switching loop adds to the bus voltage, and the capacitor sees that peak too.
The capacitance requirement comes from the allowed voltage dip during a load step. For a three-phase inverter, the minimum capacitance is approximately C = P_out ÷ (2 × f_line × V_bus × dV). Doubling the capacitance halves the dip on paper, but it also doubles the footprint, weight, and cost.
The operating constraint is almost always ripple current. The ripple current the capacitor must handle reaches its maximum at 50% duty cycle, and it flows through the ESR, the equivalent series resistance, where it becomes heat. In a 500 kW inverter at 1,200 V DC the switching-frequency ripple can reach 60 to 100 A.
The film capacitor's maximum hot-spot temperature is limited to 105 °C for polypropylene film, with self-heating of at most 20 °C, so the heatsink temperature should not exceed 85 °C. A capacitor that overheats from ripple current fails long before its capacitance matters, which is why the ripple-current rating, not the capacitance value, is the number to check first on the datasheet.
And the capacitor must sit close to the power stage. The DC-link film capacitor is mounted directly to the switches with very low ESL and ESR values, and an overlapping busbar keeps the stray inductance as low as possible. The vicinity of the capacitor to the power module is one essential target, because the inductance between the capacitor and the switches is part of the commutation loop, and that loop's inductance is what produces the overshoot in the next section.
Layout: the inductance you did not put on the schematic
Conductors that appear to be simple connections on a schematic have inductance in the physical circuit. During fast switching, this parasitic inductance can generate voltage overshoot, ringing, electromagnetic interference, and additional switching loss. It cannot be eliminated completely. It can be understood, estimated, and minimised through module selection, mechanical design, PCB layout, busbar design, and gate-drive engineering.
The voltage generated across an inductance is V = L × di/dt. Consider a switching loop with 40 nH of total parasitic inductance. If the current changes at 5,000 A/µs, the inductive voltage is 200 V, and in a high-voltage converter, an extra 200 V can consume a significant part of the semiconductor's voltage margin. The peak device voltage is approximately Vpeak = VDC + L × di/dt, and excessive overshoot can approach or exceed the module's rated blocking voltage, reducing reliability or causing immediate failure.
The most important value is usually the total inductance of the current loop, not the inductance of one conductor considered in isolation. Minimising the physical area enclosed by the commutation loop is the single most effective strategy for reducing stray inductance. That is a layout rule, and it is the reason Article #5's PCB guidance matters to a power stage: the DC-link capacitors go next to the bridge, the high-current loop is kept short and wide, and the gate-drive loop is kept separate from the power loop, because the gate loop's inductance slows the switching, and the power loop's inductance kills the switch.
Two more layout consequences deserve naming. The negative switch-node spike, parasitic inductance driving the node 20 V below ground, is what latches up high-side drivers, and it is a layout problem, not a driver problem. And the ringing that follows a fast edge is the loop inductance resonating with the device capacitances; the high di/dt of the high-frequency loop current and the energy stored in the parasitic inductance can produce phase-voltage overshoot that exceeds the absolute maximum ratings of the gate-drive circuitry, and the circuit will ring for many cycles if the overshoot is present.
The gate resistor is the dial for this: a larger resistor slows the edge, cuts the overshoot and the emissions, the subject of Article #19, and pays for it in switching loss. There is no setting that is free.
Heat: the resistance that climbs as it warms
Rds(on) increases with junction temperature, typically by 40–80% between 25 °C and 125–150 °C. Over the commercial temperature range from 25 to 150 °C, the resistance increases two to three times. The ratio K = Rds(on) at 150 °C divided by Rds(on) at 25 °C runs from about 1.5 for low-voltage transistors to about 2.9 for high-voltage ones. The maximum junction temperature is 150 °C for most parts, 175 °C for some.
That is a feedback loop, and it is the quietest way a driver dies. The MOSFET heats up, its resistance rises, its conduction loss rises as the square of the current times that resistance, and it heats up more. On its own the loop is stable, the resistance rise is not steep enough to run away in a properly cooled part, but it means the loss calculated at 25 °C is not the loss the board sees.
The demo below shows the difference: the same six MOSFETs dissipate 0.30 W of conduction loss at 25 °C and 0.48 W at 125 °C. The design must be checked at the hot resistance, not the room-temperature one, and the thermal resistance from junction to ambient, RθJA, must be the value for the copper the board actually has, not the manufacturer's test pad. On a printed circuit board with minimal traces the temperature will be higher.
Protection: the driver that saves itself
A serious gate driver does not only drive. It watches. The DRV8323 and DRV8305 both sense the VDS of the high-side and low-side MOSFETs to protect them from overcurrent, if the voltage across a conducting MOSFET rises past the threshold that means its current is too high, the driver shuts the gates down.
Both have thermal warning and shutdown, and undervoltage lockout on the gate supply: if the supply falls below the threshold, all gates go low rather than drive the MOSFETs weakly, because a weakly driven MOSFET is a resistor, and a resistor in a power stage is a heater.
The protection chain is only as fast as its slowest link. The DRV8303's datasheet gives the scale: 200 ns from an error event to all gates low, and 200 ns from the event to the fault pin going low. The microcontroller's software fault handler, running on a millisecond timescale, is not the protection, it is the report. The protection lives in the driver, in hardware, in hundreds of nanoseconds. That is the difference between a board that survives a stalled motor and a board that does not.
the Example Labs shoulder drive, sized on paper
The bench is the UR5e-class cobot from Article #10, and the joint is the shoulder from Article #20: 126.8 N·m at the joint, through a 100:1 planetary gearbox at 93% efficiency, needing 1.36 N·m at the motor at 1,500 RPM. The motor's mechanical power is 199 W, and at 90% efficiency the drive draws 221 W from the 48 V pack. The drive stage below is the 48 V, 20 kHz, 10 A peak design that powers that motor. Every number was run in Python before writing.
Step 1: the gate: what the driver must push
The MOSFET is a 60 V device with Qg = 30 nC and Qgd = 10 nC, a plateau voltage of 4.5 V, driven from 12 V through a 2 Ω driver and a 10 Ω gate resistor, a total of 12 Ω:
- Peak gate current: 12 V ÷ 12 Ω = 1.00 A, fifty times the 20 mA a microcontroller GPIO pin can source, which is why the microcontroller does not drive the gate itself
- Plateau current, turn-on: (12 − 4.5) ÷ 12 = 0.625 A; turn-off: 4.5 ÷ 12 = 0.375 A, turn-off is slower, because the driver only has the plateau voltage to work with
- Turn-on transition: 10 nC ÷ 0.625 A = 16.0 ns; turn-off: 10 nC ÷ 0.375 A = 26.7 ns
- Gate drive power: 30 nC × 12 V × 20 kHz = 7.2 mW per device, 43 mW for all six, the cheapest power in the whole design, and the reason the gate resistor is a free dial only in the sense that it costs milliwatts to turn it
Step 2: the losses: what the six switches actually dissipate
With a 48 V bus, 10 A peak phase current (7.07 A RMS), and Rds(on) = 2 mΩ:
- Switching loss per device: ½ × 48 V × 10 A × (16.0 + 26.7) ns × 20 kHz = 0.205 W; all six devices, 1.23 W
- Conduction loss per device: 7.07² × 0.002 Ω × 50% duty = 0.050 W; all six, 0.30 W at 25 °C, and 0.48 W at 125 °C, with Rds(on) 1.6× higher
- Total stage loss: 1.23 + 0.30 + 0.04 = 1.57 W on 221 W in, the drive stage's own efficiency is 99.3%, and the switching loss is four times the conduction loss, which is what 20 kHz does to a 48 V design
Step 3: the dead time: what the pause costs
At 48 V and 20 kHz, the voltage error per leg is t_dead × f_s × V_dc:
- 50 ns dead time: 0.048 V per leg, 1.0% of a 5 V commanded phase voltage
- 100 ns: 0.096 V, 1.9%
- 200 ns: 0.192 V, 3.8%
- 400 ns: 0.384 V, 7.7%
Two hundred nanoseconds is 0.4% of the 50 µs switching period, and it still costs nearly 4% of the command at low speed. The dead time should be as short as the MOSFETs' turn-off allows, the DRV8323's 50 ns setting exists for exactly this reason, and the error that remains is compensated in software, using the current's sign from the sensing in Step 6.
Step 4: the shoot-through that dead time prevents
If both devices in one half-bridge conduct together, the 48 V bus is shorted through two on-resistances of 2 mΩ each:
- 48 V ÷ 0.004 Ω = 12,000 A in theory, limited in practice by the loop inductance and the supply, but the order of magnitude is the point: the current is not a fault current, it is a dead short, and the devices fail in microseconds, before any fuse or software can react
That number is why the dead time exists, and why the driver's automatic handshaking, the DRV8305's interlock that prevents both devices of a half-bridge from switching on at the same time, is a hardware feature, not a software convention.
Step 5
The current changes at 10 A in 16 ns, 625 A/µs. The overshoot is L × di/dt, and it adds to the bus voltage across the switch that just turned off:
- 10 nH loop: 6.2 V of overshoot, Vpeak = 54.2 V, a 60 V MOSFET survives
- 40 nH loop: 25.0 V of overshoot, Vpeak = 73.0 V, a 60 V MOSFET dies, and the failure looks like a bad part
- 100 nH loop: 62.5 V of overshoot, Vpeak = 110.5 V, an 80 V MOSFET dies too
The 60 V device is the right choice for a 48 V bus only if the loop inductance is under about 20 nH. The voltage rating is not decided by the bus. It is decided by the bus plus the layout.
A board that works on the bench and dies in the field is very often a board whose loop inductance was fine at the test current and fatal at the real one, or whose replacement MOSFET switched faster than the one it replaced, and the same layout produced more overshoot.
Step 6: the sensing: what the FOC loop can actually see
Three 1 mΩ low-side shunts, one per phase, with the DRV8323's current-sense amplifiers at 20 V/V gain:
- At 10 A peak: 10 mV across the shunt, 200 mV to the ADC, and 0.10 W dissipated in each shunt
- The sampling window: at 50% duty the low-side is on for 25 µs, and the measurement needs about 1.7 µs (dead time, blanking, settling, ADC acquisition), it fits easily. At 95% duty the window is 2.5 µs, it still fits, barely. At 99% duty the window is 0.5 µs, and the measurement needs 1.7 µs, it does not fit, and the FOC loop goes blind at the top of the duty range. That is the moment to switch to in-line shunts, or to cap the duty cycle below 99%, and it is a decision to make on paper, not a discovery to make in the field
Step 7
- DC-link ripple current: worst case at 50% duty, about 5.0 A RMS on this design (the article's estimate, half the 10 A peak phase current), the capacitor's ripple-current rating must clear that number, and its voltage rating must clear the 48 V bus plus the overshoot from Step 5, with 10–20% margin
- Bootstrap capacitor: the charge per high-side on-time is the gate charge (30 nC) plus the level-shifter charge (5 nC, the article's assumption) plus the floating section's quiescent current over the on-time (100 µA × 47.5 µs at 95% duty = 4.75 nC), 39.7 nC in total. At a 1 V allowed droop, C ≥ 40 nF, and ceramic capacitors lose half or more of their capacitance under DC bias, so the nominal value placed is 100 nF, not 40 nF
- And the limit that no capacitor fixes: at 100% duty the low-side never conducts, the bootstrap capacitor never recharges, and the high-side switch turns off on its own. If the application needs 100% duty, the driver needs a charge pump, the DRV8323 and DRV8305 both have one, not a bigger bootstrap capacitor
The decision test
Five questions, in order. If you cannot answer the first, the rest do not matter.
- What is the peak gate current your driver must supply, and what does your microcontroller's pin give? If the answer is 1 A and 20 mA, the microcontroller does not drive the gates. The driver is not optional, and it is not a cost to save.
- What is your dead time, and what voltage error does it cost at your lowest commanded voltage? If the error is more than a few percent of the command, the dead time is too long for the MOSFETs you chose, or the MOSFETs are too slow for the dead time you set.
- What is your loop inductance, and what does it add to the bus voltage at your real di/dt? If the peak exceeds the MOSFET's rating, the rating is wrong or the layout is wrong, and the layout is usually the cheaper one to fix.
- Where do your shunts sit, and can the measurement finish inside the low-side window at your maximum duty cycle? If the window closes before the ADC finishes, the FOC loop goes blind at the top of the range, and the first symptom is a motor that misbehaves only at full speed.
- What happens at 100% duty cycle? If the answer is "the bootstrap capacitor runs dry and the high side turns off," the driver needs a charge pump, or the duty cycle needs a cap, decided on paper, not discovered in the field.
Objections
"But the MOSFETs are rated for 60 V and the bus is only 48 V, that's plenty of margin." The margin is not bus-to-rating. It is bus-plus-overshoot-to-rating, and the overshoot is decided by the layout. A 40 nH loop at 625 A/µs adds 25 V to a 48 V bus, and a 60 V MOSFET dies at 73 V. The rating is decided by the loop, not the bus.
"But it works on the bench." The bench is the kindest environment a driver will ever see: a short cable, a cold board, a light load, a duty cycle that never reaches the top. The field is a stalled motor at full current, a hot enclosure, a long cable, and a duty cycle at 99%. The bench proves the schematic. The field proves the layout, the thermal design, and the protection chain.
"But the driver IC costs more than three discrete half-bridge drivers." The driver IC is the dead time, the level shifting, the charge pump, the current-sense amplifiers, the VDS overcurrent sensing, the thermal shutdown, and the undervoltage lockout, in one part, tested together, with a datasheet that states the timing. The discrete version is six drivers, three bootstrap circuits, three amplifiers, and a protection scheme you have to design, test, and guarantee yourself. The IC is cheaper by the time the second board is built.
"But GaN switches an order of magnitude faster, so it must be better." Faster switching is lower switching loss and higher overshoot, in the same step. A layout that performed acceptably with a silicon MOSFET may produce excessive overshoot or ringing after a GaN device is substituted, because the di/dt is higher across the same loop inductance. Wide bandgap is a better switch and a harder layout problem, and the two arrive together.
Sources
- United Motion Inc., "What Are the Fundamental Principles of a BLDC Motor Driver Circuit?"
What we used
A typical BLDC driver contains five functional sections (DC power supply and DC-link capacitor; three-phase MOSFET inverter; gate-driver circuit; microcontroller or motor-control IC; position, current, voltage, and temperature feedback); the three-phase inverter is six MOSFETs; the MCU normally should not drive the MOSFET gates directly in a serious power stage; current feedback is used for overload detection, torque control, current limiting, and fault protection
- UMA Technology, "Implementing a 3-Phase Brushless Direct Current Motor Drive"
What we used
A 3-phase bridge made from six MOSFETs or IGBTs connects the bus to motor phases U, V, and W, each phase driven by one high-side and one low-side switch; DC link, inverter, gate driver, rotor position feedback, and controller as the core blocks
- Wevolver, "BLDC Motor Controller: Comprehensive Design Guide for Engineers"
What we used
The controller's subsystems (microcontroller firmware, gate drivers, three-phase inverter, current/voltage sensing, protective circuits); six-step, sinusoidal, and FOC commutation; the gate driver amplifies MCU logic signals to drive the MOSFET or IGBT power switches
- Microchip, AN5324, "High-Power BLDC Motor Driver Reference Design"
What we used
18–85 V input, 10 kW peak, 600 A peak-to-peak phase currents; 150 A continuous, 300 A peak for 120 s; six 250 µΩ low-side shunts; eighteen MOSFETs at 1.7 mΩ Rds(on); PWM 8–20 kHz, typical 16 kHz; MCP9700 temperature sensor on the power MOSFETs
- Texas Instruments, DRV832x datasheet SLVSDJ3D (Rev. D, March 2022)
What we used
DRV8323: 6 to 60 V operating voltage; drives 3 high-side and 3 low-side N-channel MOSFETs; adjustable slew-rate control; 10 mA to 1 A peak source current, 20 mA to 2 A peak sink; supports 100% PWM duty cycle; adjustable dead time 50/100/200/400 ns over SPI (hardware version 100 ns); three integrated low-side current-sense amplifiers, adjustable gain 5/10/20/40 V/V; supports 1.8 V, 3.3 V, and 5 V logic inputs; integrated charge pump for the high side
- Texas Instruments, DRV8305 datasheet SLVSCX2B (Rev. B, February 2016)
What we used
4.4 to 45 V operating voltage; 1.25 A and 1 A peak gate drive currents; programmable high- and low-side slew-rate control; charge-pump gate driver for 100% duty cycle; three integrated current-shunt amplifiers with variable gain and adjustable offset; integrated 50 mA LDO (3.3 V or 5 V); 3-PWM or 6-PWM input control up to 200 kHz; automatic handshaking when switching to prevent shoot-through; VDS sensing of both high-side and low-side MOSFETs for overcurrent protection
- Texas Instruments, DRV8303 datasheet SLOS846C
What we used
6 to 60 V; adjustable dead-time control with an external resistor on the DTC pin, 50 ns minimum; two current-shunt amplifiers; gate drive rise and fall times 25 ns at 1 nF load; 200 ns from an error event to all gates low
- Texas Instruments, DRV8328 datasheet SLVSFF3C
What we used
Dead-time insertion to prevent both external MOSFETs of each half-bridge from switching on at the same time; cross-conduction prevention when high and low side are commanded on together
- Mouser, "Powering Electric Motor Drives"
What we used
MOSFETs can potentially switch at frequencies up to 100 kHz but are more often used in the tens of kHz; IGBTs operate at lower frequencies of around 20 kHz and need an external anti-parallel diode; Rds(on) and gate capacitance are the most important secondary MOSFET parameters; the higher the switching frequency, the greater the losses and the lower the efficiency
- Avnet, "Selecting Wide Bandgap Devices for Motor Control Applications"
What we used
IGBT switching frequencies for motor controllers are typically 10–20 kHz; Si MOSFETs are often used up to a few kW and to around 1,200 V; SiC is superior to silicon in thermal conductivity; GaN switches an order of magnitude faster than silicon
- JTDZ Tech, "MOSFET or IGBT? Selecting Power Switching Devices for Real Loads"
What we used
MOSFET conduction loss is I² × Rds(on); IGBT conduction loss is roughly I × VCE(sat); IGBTs dominate above roughly 600–1,200 V and hundreds of amps; practical ceilings around 20–50 kHz for standard IGBTs; the voltage/frequency grid (<200 V MOSFET/GaN, 200–600 V SiC vs superjunction, 600–1,200 V IGBT or SiC, 1,200–6,500 V IGBT)
- Engineerz Corner, "Si vs SiC vs GaN Power Semiconductors"
What we used
SiC devices rated to roughly 1,200 V+ with much lower switching losses than silicon IGBTs; GaN switching frequencies well into the megahertz range with no reverse-recovery losses, strongest below roughly 650 V; silicon IGBTs the cost-effective choice for high-power, lower-frequency work
- A2Z Tools, "MOSFET Gate Resistor Calculator"
What we used
The Miller charge Qgd sets the switching time, not the total Qg; plateau current on turn-on is (Vdrive − Vplateau) ÷ Rtotal, on turn-off Vplateau ÷ Rtotal; gate drive power is Qg × Vdrive × frequency; worked example, 48 V half-bridge at 100 kHz: Qg 30 nC, Qgd 10 nC, 12 V drive, 4.5 V plateau, 2 Ω driver plus 10 Ω gate resistor, peak gate current 1 A, plateau 0.625 A, turn-on 16 ns, turn-off 26.7 ns, switching loss 0.512 W, conduction 0.25 W, gate drive 36 mW
- Firgelli, "MOSFET Gate Driver Interactive Calculator"
What we used
Ipeak = Vdriver ÷ (Rg + Rdriver); Iavg = Qg × fsw; trise ≈ Qg ÷ Igate; tmiller = Qgd ÷ Igate; Egate = Qg × Vgs; Pgate = Qg × Vgs × fsw
- Power Electronics Hub, "Bootstrap Capacitor Calculator"
What we used
The high-side driver is powered from a capacitor that floats on the switch node, refilled through the bootstrap diode only while the low-side switch conducts; C ≥ Q ÷ ΔV; ceramic capacitors lose capacitance under DC bias, often by half or more; a bigger capacitor does not help when the low-side time is short, only a smaller resistance or a shorter maximum duty does
- PCBSync, "Gate Driver IC Selection: Bootstrap, Isolated & High-Side"
What we used
A gate driver delivers the amps of peak current needed to switch a power FET fast, sized from Ipeak ≈ Qg ÷ rise time; bootstrap cannot run at ~100% duty, it needs regular low-side conduction to refresh; the negative switch-node spike that latches up high-side drivers is a layout problem, with parasitic inductance driving the node 20 V below ground
- onsemi, AN-6076, "Design and Application Guide of Bootstrap Circuit for High-Voltage Gate-Drive IC"
What we used
The bootstrap power supply is simple and low cost, but the on-time duty cycle is limited by the requirement to refresh the charge in the bootstrap capacitor, and serious problems occur when negative voltage is presented at the source of the switching device
- Texas Instruments, SLUA887A, "Bootstrap Circuitry Selection for Half-Bridge Configurations"
What we used
The bootstrap circuit is a capacitor, a diode, a resistor, and a bypass capacitor; it generates high-side bias referenced to the switch node
- Texas Instruments, Analog Design Journal SLYT762, "Selecting Amplifiers for Shunt-Based Current Sensing in 3-Phase Motor Drives"
What we used
The three shunt placements (DC-link single shunt, two or three low-side shunts, two or three in-line motor shunts); in-line shunts give direct, continuous phase-current sense, allow averaging over one PWM cycle, and can detect phase-to-phase and phase-to-ground shorts and shoot-through, but need high common-mode input voltage and CMRR; low-side shunts are indirect and discontinuous, can only be measured when the low-side switch is on, and cannot be averaged over one PWM period; phase currents can be as high as 100 A in inverters from 12 to 60 V DC
- TI E2E forum, "TMS320F2806: Inline vs Low Side Current Sensing for FOC Algorithms"
What we used
In-line sensing gives continuous knowledge of the current feedback, while low-side shunts give feedback only when the low-side switch is on; at high duty cycle the low-side on-time can effectively become zero; the ADC start-of-conversion trigger must be synchronised with the PWM for low-side sensing
- GaN Servo Drives, "Current Sensing Architectures in High-Speed GaN Servo Drives"
What we used
Choose low-side shunts when the supplier demonstrates valid sampling windows throughout the required speed, load, and modulation range; consider in-line sensing when low-side sampling windows restrict operation; the conservative timing budget is t_min = t_dead + t_blanking + t_settle + t_acquisition; neither topology nor sensor isolation establishes Safe Torque Off capability
- EE Power, "Calculating DC-Link Capacitance for xEV Powertrains"
What we used
The DC-link capacitor absorbs current ripple generated by the fast-switching three-phase inverter; the ripple current reaches a maximum at 50% duty cycle; the film capacitor's maximum hot-spot temperature is 105 °C with self-heating of at most 20 °C; the capacitor must be mounted directly to the switches with very low ESL and ESR
- AT Capacitor, "Selecting DC Filter Capacitors for High-Power Inverter Systems"
What we used
The voltage rating must be chosen for the worst-case transient, at least 10–20% above the steady-state bus; the operating constraint is almost always ripple current; in a 500 kW inverter at 1,200 V DC the switching-frequency ripple can reach 60–100 A
- Power Electronics Hub, "Dead-Time Voltage Distortion in Three-Phase Inverters"
What we used
The average error of one leg is Δv = −sign(i) × t_dead × f_s × V_dc; a 540 V, 8 kHz, 3 µs drive loses about 13 V per leg, and at 5 Hz the same 13 V is a large fraction of the commanded voltage; the 5th and 7th harmonic currents combine into torque ripple at six times the electrical frequency; a SiC case at 800 V, 20 kHz, 200 ns loses 3.2 V, about 1.1% of the fundamental at m = 0.9
- Analog Devices, AN331-50, "Compensating the Dead Time of Voltage Inverters"
What we used
The dead time introduces small voltage errors sufficient to produce distorted motor currents, oscillations of the motor torque, and loss of controllability; compensation is by feed-forward of the dead-time voltage into the reference
- MDPI Energies, "Dead-Time Inverter Voltage Drop in Low-End Sensorless FOC Motor Drives" (2024)
What we used
Dead-time inverter voltage drops contribute to suboptimal controllability at low motor speeds; voltage drops are typically modelled as an average voltage loss over the switching period and compensated directly within the abc-reference frame
- Fuji Electric, "Understanding Parasitic Inductance in Power Modules"
What we used
V = L × di/dt; a 40 nH loop at 5,000 A/µs produces 200 V of inductive voltage; Vpeak = VDC + L × di/dt; minimising the physical area enclosed by the commutation loop is the single most effective strategy; a layout that performed acceptably with an older IGBT may produce excessive overshoot after a faster device is substituted
- Power Electronic Tips, "How Do Parasitic Inductances Affect Switching Performance?"
What we used
Stray inductance causes voltage spikes, ringing, and higher switching losses; the overshoot adds to the normal operating voltage and can exceed the transistor's breakdown voltage rating
- Allegro, AN296215, "Motor Drive Discrete MOSFET Bridge Circuit Design and Layout"
What we used
The high di/dt of the high-frequency loop current and the energy stored in the parasitic inductance can result in phase-voltage overshoot exceeding the absolute maximum ratings of the gate-drive circuitry; the parasitic inductance and capacitance form a high-Q resonant circuit that can ring for many cycles
- EE Times, "Developing Analytical Equations for Determining Power MOSFET Switching Transients"
What we used
The gate-charge waveform divides into Qgs, Qgd (the Miller plateau), and Qg; calculated switching transients will always be shorter than those actually achieved, so maximum datasheet parameters should always be used
- MST Tutorial, "MOSFET as a Switch"
What we used
Rds(on) increases with junction temperature, typically by 40–80% between 25 °C and 125–150 °C; the AO3400A datasheet gives 26.5 mΩ maximum at 25 °C and 38 mΩ at 125 °C (factor 1.43); maximum junction temperature 175 °C for the TO-220 parts, 150 °C for the AO3400A
- Electronic Design, "Analyze MOSFET Parameter Shifts Near Maximum Operating Temperature"
What we used
K = Rds(on) at 150 °C divided by Rds(on) at 25 °C runs from about 1.5 for low-voltage transistors to about 2.9 for high-voltage ones; the IPB60R099 datasheet gives 0.09 Ω at 25 °C and 0.23 Ω at 150 °C
- Article #20, "Choosing the Motor: Torque Decides, Not RPM" (Senova, 2026-11-19)
What we used
The shoulder's 126.8 N·m, the 1.36 N·m motor at 1,500 RPM through a 100:1 planetary at 93% efficiency, the 199 W mechanical power and 221 W electrical draw, the 48 V / 100 Ah (4,800 Wh) pack, and the FOC loops that live in the microcontroller
- Article #19, "EMC/EMI: The Board That Fails in the Chamber, Not on the Bench" (Senova, 2026-11-19)
What we used
The gate resistor as the dial between overshoot and emissions, and the 16.1 dB a four-layer board buys back
- Article #5, "PCB Design for Manufacturing" (Senova, 2026-11-19)
What we used
The board the power stage lives on, and the placement rules the commutation loop depends on