Doubling the PWM frequency added about ten degrees at the case of a GaN half bridge; bolting on a heatsink took thirty off. The thermal path, not the switch, sets how fast you get to run. Every actuator designer eventually meets the same tradeoff. Raising the PWM frequency tightens control bandwidth and shrinks the magnetics, but it also raises switching loss and pushes the device hotter. The question is how much thermal room you actually have, and one vendor bench gives a clean answer for GaN.
What the bench shows
Innoscience's demo manual for its 48 V, 22 A motor drive on the ISG3204LA, a 100 V half bridge with integrated driver, reports case temperature against phase current at two switching frequencies with a heatsink, and at 20 kHz without one under natural cooling [1]. At 20 kHz and 20 A rms with the heatsink fitted, the case measured about 81.1 °C. Raising the switching frequency to 40 kHz lifted it by about 10 °C, to 90.9 °C on the chart [1]. Without the heatsink, the same 20 kHz, 20 A point sat at 112.8 °C [1].
That is a shallow slope against frequency and a steep one against construction. The heatsink was a 70 mm by 35 mm by 25 mm block bonded to the underside of the PCB through a thermal interface material with a conductivity of 4 W/mK, and Innoscience states the result as roughly 30 °C lower than the bare board [1]. Put the two numbers side by side and the arithmetic is ours: one construction choice bought back the equivalent of three frequency doublings.
The thermal path is the real budget
The reason the slope is shallow is that the GaN device starts with low switching loss. On the same bench at 20 kHz, the ISG3204LA lost 12.3 W against 16.3 W for a similar on resistance silicon MOSFET at 20 A rms, and 1.2 W against 2.6 W at 5 A rms [1]. TI's GaN reference design tells the same story from the other direction: measured at 40 kHz, 60 kHz and 80 kHz, the board losses were almost identical at 40 kHz and 80 kHz [2]. Our reading of both data sets is a device whose loss is dominated by conduction once the frequency is in the tens of kilohertz, which is why the case temperature moves so little when the frequency doubles.
The silicon comparison is what makes the point sharp. Innoscience measured the GaN device about 20 °C cooler than the silicon part at 20 kHz without a heatsink [1]. We expect that gap to open with frequency, because the silicon loss is the one with the larger switching term in it, though the manual does not give the silicon curve at 40 kHz. In our reading, the frequency budget of a silicon stage is set by the switch; the frequency budget of a GaN stage is set by the heatsink.
Why GaN gets to push frequency
Because GaN starts with far lower switching loss, the same thermal path carries it to a higher frequency before the case reaches its limit. For a robot joint that needs control bandwidth, that headroom is the product. You spend it on a faster current loop, on smaller magnetics, or simply on a cooler, longer lived drive. TI's joint drive work puts the current loop at 8 kHz to 32 kHz with the switching frequency typically equal to it [2], so a stage that shrugs off 40 kHz is already past the loop rate a humanoid joint asks for.
What we would push on is the 81.1 °C itself. That is the case with a 70 by 35 by 25 mm heatsink under natural convection, which is more metal than a joint housing gives you. A joint drive gets the housing as its heatsink and the motor as its neighbour, and the number that matters is the case temperature at 40 kHz through the housing's actual thermal resistance. Nobody publishes that, because it is a system number, and it is the one we intend to measure on ulna, the board of our own bring up notes [3].
References
- Innoscience, DB043 INNDMD48V22A1 48 V 22 A Motor Drive Featuring ISG3204LA Demo Manual, rev. 1.0, Chinese edition, Figures 9, 10 and 11.
- Texas Instruments, Application of GaN FET in Humanoid Robots, SNOAAB3, January 2025.
- Aleph-B, ulna-60-30 BLDC controller test report, v1.0, 22 June 2026, internal, the board the measurement above is planned on.