Novomorphic

Article

GaN is not the product

System architecture decides performance.

8 min read

Most wide-bandgap disappointment is not a device problem. It is a switching, layout, sensing and control problem being solved by four separate teams.

Article by Novomorphic engineeringReviewed by Technical review, Novomorphic

GAN HALF BRIDGEFast switches, set by the systemThe device is fast. The system decides whether that speed is usable.DEAD TIMEBOARD LAYOUTONE PHASE LEGGROUND PLANEDC link +DC link 0Phase nodeLink capacitorsHigh sideGaN switchLow sideGaN switchCurrent sensePhase outControllerFPGA, FOCGate driver, highGate driver, lowHigh side gateLow side gatePhase nodeDead timeTime
Fig. 1: GaN power stage

Summary

The device delivers exactly what the system around it allows

A gallium nitride transistor switches faster than silicon, has no reverse-recovery charge, and has a fraction of the gate charge. On paper, a GaN power stage is smaller, cooler and more efficient. On the bench, a first GaN design very often measures worse than the silicon one it was meant to replace, and the explanation offered is usually “GaN is immature”. It is not. The device did what it was asked. The gate drive, the layout, the dead time, the sensing and the control loop asked it the wrong things, and they asked separately because four different people owned them.

This article explains where the performance goes, with the figures from the device makers’ own application notes, how we measure it rather than simulate it, and why we organise a power-stage programme as one body of work from the control law to the copper.

The gate

A 1.5 V margin and a nanohenry budget

An enhancement-mode GaN FET has an absolute maximum gate voltage of about plus 6 V and minus 5 V, against plus or minus 20 V for a silicon MOSFET. The recommended drive is around 4.5 V, which leaves a 1.5 V margin to the maximum. Any ringing on the gate eats that margin, and ringing is set by the gate loop inductance against the gate capacitance. EPC’s guidance is that the gate loop inductance should stay below a quarter of the squared gate resistance times the gate capacitance for near-critical damping. In practice that is a loop of a few nanohenries, which is a layout number, not a circuit number.

The second gate problem is common-source inductance: the inductance shared between the power current path and the gate return. Every nanohenry of it converts the device’s own di/dt into a gate voltage that fights the driver. EPC is blunt that this matters more for GaN than for MOSFETs because the di/dt and dv/dt are higher. The third is dv/dt-induced turn-on: the Miller capacitance pumps charge into the gate when the opposite device switches, and the maximum tolerable dv/dt is set by the threshold voltage divided by the pull-down impedance times the gate-drain capacitance. Drivers with better than 50 V/ns capability exist. Whether the layout lets them deliver it is another matter.

Dead time

GaN has no body diode, so dead time costs more and must be shorter

In a silicon half-bridge, current during dead time flows through the body diode at about a volt. A GaN HEMT has no PN junction; reverse current flows through the same channel, and the channel only opens once the source-drain voltage exceeds the threshold voltage plus whatever negative gate voltage the driver is holding. Infineon’s own description is that GaN shows a higher drop in third-quadrant operation than silicon, and that the drop increases with negative gate bias and with temperature. Hold the gate at minus 3 V for safety against dv/dt turn-on, and dead time costs several volts of drop at the full phase current.

Texas Instruments puts the arithmetic plainly: dead-time conduction loss is proportional to switching frequency, load current, the voltage drop and the dead-time duration. Every one of those is set by a different team. Frequency by the control engineer, current by the application, the drop by the gate drive designer, and the duration by whoever set a conservative constant in the firmware at commissioning and never revisited it.

The device makers have responded in hardware. TI’s integrated three-phase GaN drive stage uses adaptive dead time with a maximum under 200 ns and a propagation delay under 200 ns, against the microsecond or more that conventional discrete solutions need, and reports better than 99% inverter efficiency on a 250 W fan drive at 20 kHz with the heat sink removed. The same note offers slew-rate control down to 5 V/ns as an explicit trade between motor winding insulation and switching loss. That trade is the point of this article: it cannot be made by the power engineer alone.

Four teams, one node

Where the performance is lost between the disciplines

Dead time

Owned separately
A constant chosen for worst case, often 500 ns to 1 µs
Owned as one system
A control variable: minimised at the operating point, bounded for safety, measured

Slew rate

dv/dt at the switching node.

Owned separately
As fast as the device allows, then EMC fails
Owned as one system
Set against motor insulation, common-mode current and loss together

Layout

Gate and power loop inductance.

Owned separately
Schematic first, layout later, ringing discovered on the bench
Owned as one system
Loop inductance budgeted in nanohenries before the schematic is finished

Current sensing

Bandwidth and timing relative to the switching edge.

Owned separately
Sampled wherever the ADC trigger lands, corrupted by edge noise
Owned as one system
Sampled at a defined point in the cycle, in hardware, with the edge excluded

Thermal

Owned separately
Heat sink sized from the data sheet
Owned as one system
Loss measured at the real operating points, heat path designed with the enclosure

Reliability

Dynamic on-resistance and switching stress.

Owned separately
Static data sheet parameters
Owned as one system
Evaluated under switching conditions following JEDEC JEP180

Measure, then believe the model

The double-pulse test is where the argument ends

Switching loss cannot be read from a data sheet, because it depends on the gate drive, the layout and the load at the point of use. It is measured with a double-pulse test: the device is switched twice into an inductive load from a known bus voltage so that turn-on and turn-off energies can be extracted at a defined current. Infineon’s worked example uses 400 V and 10 A into 79.2 µH for a 650 V device. The method is standard; doing it well on GaN is not, because the edges are a few nanoseconds and probe bandwidth, deskew between voltage and current channels, and the probe’s own ground loop all distort the result.

Reliability is measured under switching too. JEDEC’s JC-70 committee published JEP180 in February 2020 precisely because the static qualification tests written for silicon do not exercise the mechanisms that age a GaN device in service, dynamic on-resistance among them. A GaN programme that qualifies its device with silicon tests has not qualified it.

Our rule is simple. A simulation whose model has not been checked against a double-pulse measurement on the real layout is a hypothesis. We instrument the first boards, extract switching energy across the operating range, and feed the measured loss map back into both the thermal design and the control strategy. The remaining efficiency is usually in the control strategy, which is the one place a power-electronics team working alone cannot reach.

How we organise the work

From the control law to the copper, one team

  1. Envelope and topology

    Voltage, current, frequency, thermal budget, form factor and cost. Whether GaN, silicon carbide or silicon is right follows from these, and the honest answer at some volumes is silicon.

  2. Budget the loops

    Gate loop and power loop inductance in nanohenries, before the schematic is finished. The layout is designed to the budget, not checked against it afterwards.

  3. Design the switching with the control

    Dead time, slew rate, protection thresholds and the ADC sampling instant are decided together with the control engineer, because each one moves the others.

  4. Measure under real conditions

    Double-pulse characterisation on the real layout, then efficiency, thermal and EMC maps across the operating range on the bench.

  5. Close the loop

    Measured loss and thermal behaviour feed back into the control strategy. Dead time becomes a variable the controller manages within bounds rather than a constant.

  6. Qualify as GaN

    Switching reliability evaluated under conditions that follow JEP180, not static silicon tests.

Where this applies

Drives, converters and always-on edge hardware

The motor drive is the clearest case, because dead time and slew rate act directly on torque ripple, acoustic noise and winding stress. The same discipline applies to a power conditioning stage inside a sealed, always-on edge device, where every watt of switching loss has to leave through the enclosure. In GRID-5G, the Secure Edge AI platform’s GaN power conditioning is designed by the same team as the deterministic datapath it feeds, inside a hard power and thermal budget for the substation environment. That design is in progress.

Treating dead time as a control variable rather than a commissioning constant is concept-stage work at Novomorphic. We do not quote a loss reduction for it. What we will say is that any supplier who quotes an efficiency improvement for your machine before seeing your data is guessing.

What this does not prove

GaN is not always the answer either

None of this means GaN is right for a given design. At high voltage and high current, silicon carbide often wins. At low frequency and low cost, silicon still does. The point is narrower: when GaN is chosen, its benefit is delivered or lost in the system around it, and a programme that separates the device, the drive, the layout and the control will lose it.

The figures quoted here are the device makers’ published values for their own parts and reference designs. They are not Novomorphic measurements and they will not transfer directly to another layout.

Sources (8)Every external fact above is cited here so it can be checked rather than believed.

Keep reading

Keep reading

All insights

Talk to the engineer, not the account manager

Tell us what you are building. We will put you in front of whoever actually knows the answer.

What happens next

1

An engineer reads it

Not a form-handler. Whoever owns the relevant discipline.

2

We come back with a technical view

Including if we think you do not need us.

3

If it looks like a fit, a short technical call

Then, if it is useful, a scoped workshop, feasibility study or evaluation.