GaN vs SiC for a 3 kW Supply: The Week We Chose Ruggedness
Archi5 min read
Contents
Field Notes: GaN vs SiC for a 3 kW PFC Stage#
The engineering team had locked the power supply specification in March, but the transistor choice stayed open until late June. We had two candidates: a 650 V gallium nitride HEMT and a 650 V silicon carbide MOSFET. Both fit the footprint. Both promised high efficiency. One week of bench testing settled the argument.
The Test Setup#
We built a half-bridge boost converter with a 400 V DC link and a 100 kHz switching frequency. The target output was 3 kW continuous with a 4 kW peak for five seconds. We used a common gate driver board with adjustable dead time and a current sense resistor in the source leg. Five samples of each transistor came from the same date code.
The test plan had four parts: double pulse switching loss, full load thermal soak, a 100 hour high temperature operating life test, and a hard short circuit test. We recorded data on a four channel oscilloscope and a thermal camera.
Switching Losses at 100 kHz#
The GaN part was clearly faster. At 10 A load and 400 V, the turn off energy was 0.16 mJ. The SiC part measured 0.29 mJ. Turn on energy was closer: 0.21 mJ for GaN and 0.26 mJ for SiC. At 100 kHz, the computed total loss difference was 1.8 W per transistor in the boost stage. The GaN converter reached 97.3 percent efficiency at full load. The SiC converter reached 96.9 percent.
That 0.4 percent gap would reduce power dissipation in a sealed enclosure. For a product rated for 50,000 hours, that difference showed up in the cost of the heatsink and the fan. The GaN design could use a smaller extruded aluminium profile and a slower fan. The SiC design needed a slightly larger profile and a 20 mm fan.
Thermal Behavior#
Efficiency was not the only metric. At full load in a 25°C room, the GaN case temperature settled at 71°C. The SiC case settled at 63°C. The GaN part was in a quad flat no-lead package with a center pad. The SiC part was in a TO-247 package with a large copper tab. The tab made mounting to the heatsink easier and reduced the interface temperature drop.
The thermal camera showed a hotspot on the GaN gate loop at 89°C during the 4 kW peak. The SiC gate loop stayed below 80°C. That narrower margin worried us, because the gate driver was only 8 mm away from a source that could reach 95°C in a fault condition.
Dead Time and Reverse Conduction#
The GaN HEMT has no body diode. In reverse conduction, it behaves like a diode with a threshold near 1.8 V. The SiC MOSFET has a real body diode with a forward drop near 2.6 V. We tuned dead time for both.
At 100 kHz, the GaN part needed 80 ns dead time. With 60 ns, we saw cross conduction and a 6 A current spike. The SiC part worked with 120 ns dead time and tolerated 200 ns without a measurable loss increase. That tolerance mattered for mass production. Gate drive timing on a production line moves with component tolerances, and a 40 ns window is hard to hold across temperature.
The 100 Hour Soak#
We ran both converters at 3 kW in a 45°C chamber for 100 hours. At 50 hours, the GaN on resistance had increased by 3 percent from the starting value. At 100 hours, it was 5 percent higher. The SiC part rose 1.2 percent and then stabilized. The GaN increase was not a failure, but it shifted the efficiency point and changed the gate drive timing slightly.
We also measured gate leakage current every 12 hours. The GaN samples stayed below 100 nA. Two SiC samples showed 50 nA. No drift beyond the meter resolution.
The Short Circuit Test#
The last test was the hard short circuit. We clamped the output and applied a 600 V bus with the gate at full turn on. The power supply current limit was disabled.
The SiC part survived 8 microseconds before the gate driver overcurrent protection turned it off. The die temperature at that moment was not directly measured, but the case sensor showed 72°C. We repeated it on three samples. All three survived.
The GaN part failed at 3 microseconds on the first sample. The failure mode was a shorted gate. We repeated with two more samples and got 2.8 and 3.2 microseconds. The vendor data sheet specified 2 microseconds minimum, so the test result was inside specification. But 3 microseconds is a narrow window. A protection circuit has to detect the fault, blank the switching noise, and turn off the gate in less than 3 microseconds. That is possible, but it leaves no room for a slow sensor or a capacitor delay.
The Decision#
At the end of the week, the engineering group voted with test data. We chose the SiC MOSFET for the production design. The 0.4 percent efficiency loss was real, but the SiC part gave us three advantages. It could survive a hard short circuit for 8 microseconds. It tolerated a wider dead time window. And its on resistance did not drift after 100 hours.
The GaN part performed better in switching loss and allowed a smaller heatsink. Those are attractive for a consumer adapter. For an industrial power supply that sits in a metal cabinet with other equipment, reliability matters more than a 4 W difference in the PFC stage.
We told the mechanical team to use the larger heatsink and the 20 mm fan. The electrical team reduced the dead time setting to 110 ns and added a small RC filter on the current sense signal. Production will start with the SiC part from one qualified fab and a second fab approve as an alternate source.
The GaN samples went back to the distributor. The field notes end with one observation: the fastest transistor is not always the right transistor. The right one is the one that still works after the next fault, the next thermal cycle, and the next batch of parts with slightly different gate charge. That was the decision, and the data put it on paper.
More articles