Switching loss is about a quarter of motor power in a silicon joint drive, GaN takes most of it back, and a bench measurement on a 48 V inverter agrees with the field report. The silicon MOSFET is the quiet ceiling on a low voltage BLDC drive. It works, but in an actuator that has to spin up fast and run cool, the switch is where the power goes. Here is what changes when you replace it with GaN, in numbers from two independent sources.
Where the power goes in silicon
The field report covers Zhiyuan, the Shanghai humanoid maker, fitting Innoscience GaN into its joint drives [1]. It states that silicon switching loss accounts for about 23 percent of the motor's total power consumption, that GaN cuts switching loss by 85 percent and shrinks power device chip area by more than 50 percent, given as general device statements, and that Innoscience's own tests show the joint drive board volume down by about 50 percent [1]. Those are a business publication's figures, not a datasheet, and we have no second source for them; we carry them as reported.
The chip area figure has a vendor side twin. TI compares a silicon MOSFET three phase reference design at 720 mm² of power device area against a GaN design at 320 mm², a reduction it states as more than 50 percent [4]. Two different companies, two different power stages, the same headline.
The same story on a bench
The bench measurement comes from Innoscience's own demo manual for a 48 V, 22 A motor drive built on the ISG3204LA, a 100 V half bridge with the gate driver in a 5 mm by 6.5 mm LGA package [2]. On an inductive load bench at 48 V and 20 kHz, with the silicon dead time set to 1 µs and the GaN dead time to 100 ns, system loss measured 2.6 W for the silicon MOSFET against 1.2 W for GaN at 5 A rms per phase, a 53.9 percent reduction, and 16.3 W against 12.3 W at 20 A rms, a 24.5 percent reduction [2].
Read the two load points together. At 5 A the loss is dominated by switching and the GaN advantage is large. At 20 A conduction loss takes over, the two devices have similar on resistance by construction, and the advantage narrows to a quarter. In our reading that is the honest shape of the GaN case at 20 kHz: it is a switching loss story, and it grows with frequency, not with current. Innoscience also measured the GaN device about 20 °C cooler than the silicon part at 20 kHz without a heatsink [2].
The mechanism underneath both reports is the one EPC lays out in its dead time white paper: a GaN FET has no body diode and no reverse recovery charge, so in EPC's calculation for 100 V eGaN against 80 V MOSFETs on a 60 V bus the hard commutation and recovery losses are about a third of the MOSFET figure, and the optimum dead time about half [3]. The 1 µs against 100 ns dead time setting in the Innoscience test is that difference made concrete.
Why it matters now
Joint motor power is climbing from around 80 W toward 200 W as robots add degrees of freedom, and a finger joint that has to reach 500 rpm within 0.1 s requires, in the report's words, a drive that supports a higher PWM frequency [1]. Higher power at higher PWM frequency is exactly the corner where the silicon switching loss share, already 23 percent at today's ratings, gets worse. For a low voltage BLDC actuator the switch is no longer a commodity choice.
What we would push on is the 24.5 percent at 20 A. That is the number a sceptical actuator engineer will quote back, because 20 A rms is closer to a working joint than 5 A is. The answer, in our view, is not to argue the number but to move the operating point: at 40 kHz or 100 kHz, where a joint drive wants to be for current loop bandwidth, our expectation is that the silicon loss climbs with frequency while the GaN loss moves far less. Innoscience's own thermal data shows the GaN case rising about 10 °C from 20 kHz to 40 kHz, with heatsink at 20 A rms [2]; the silicon curve at 40 kHz is not given. Nobody has published the silicon curve at 40 kHz on the same bench, and that is the measurement we would want.
References
- 36Kr Europe, Is Zhiyuan's Equipping with Innoscience Chips a New Inflection Point for GaN?, 15 July 2025, https://eu.36kr.com/en/p/3379671551140609, figures as reported.
- Innoscience, DB043 INNDMD48V22A1 48 V 22 A Motor Drive Featuring ISG3204LA Demo Manual, rev. 1.0, Chinese edition, Figures 9, 10 and 11.
- J. Strydom and D. Reusch, Dead-Time Optimization for Maximum Efficiency, EPC white paper WP012.
- Texas Instruments, Application of GaN FET in Humanoid Robots, SNOAAB3, January 2025.