AI Server Power Supplies: GaN Adoption Criteria

Whether to adopt GaN for AI server power supply design is not solely decided by its status as a next-generation technology. It requires confirming whether GaN offers a clear advantage over SiC or Si across three axes: switching frequency, power density, and thermal design.

Conditions for GaN Adoption in AI Server Power Supplies

The required specifications for power supply units (PSUs) for AI servers differ from those for conventional data centers. Servers equipped with NVIDIA's H100 or B200 can exceed 100 kW of power consumption per rack, making power density and conversion efficiency directly linked to the overall system's thermal design. GaN (gallium nitride) is attracting attention in this environment due to its material advantages of high switching frequency and low switching loss.

Because GaN devices employ a lateral structure, issues specific to vertical devices like short-circuit withstand time in SiC are less likely to be a primary design concern. On the other hand, AI server power supplies are increasingly adopting configurations such as 48V to 12V conversion or direct 48V to 1V conversion (Direct-to-chip), where GaN characteristics are best utilized within a voltage range below 650V. This suggests a natural segmentation from the high-voltage, high-current inverter domain where SiC is dominant.

So, what are the specific criteria for adopting GaN? The first point to check is the operating frequency. GaN exhibits significantly lower switching losses compared to Si, allowing for high efficiency even in the hundreds of kHz to several MHz range. This high-frequency operation enables the miniaturization of passive components (inductors, capacitors), thereby reducing the overall PSU volume. As AI server rack density increases, the value of this volume reduction becomes more significant.

Design Differences in Short-Circuit Protection—How is it Different from SiC?

While GaN and SiC are sometimes discussed in the same context for AI server power supplies, the nature of design challenges differs from a short-circuit protection perspective. When selecting SiC MOSFETs, short-circuit withstand time (SCWT, Tsc) is always a key consideration. This indicates the time until the device is destroyed when a load short occurs, and it functions as a grace period for the protection circuit to operate.

SiC devices have smaller die sizes and higher current densities, resulting in faster temperature rise during short circuits compared to Si devices. Microchip's 700V/1200V SiC MOSFETs specify a short-circuit withstand time of typ. 3μs under certain conditions in their datasheets. This figure implies that the protection circuit must turn off the device within 3μs.

GaN devices exhibit different behavior due to their lateral structure. While short-circuit withstand time is not an issue that can be entirely ignored, the design approach for overcurrent protection changes for the relatively low-voltage and low-current applications adopted by AI server power supplies. Regardless of which is chosen, the response speed of the protection circuit remains inseparable from device selection.

DESAT (desaturation) detection is widely used for SiC short-circuit protection. This mechanism monitors the drain-source voltage (VDS) in the on-state and turns off the power transistor upon detecting overcurrent. It is often integrated into gate driver ICs. GaN gate drivers also incorporate overcurrent protection functions, but the balance between response speed and malfunction prevention (blanking time setting) varies depending on the design.

How to Interpret Power Density Metrics

A recurring metric in the development competition for AI server power supplies is power density (W/in³ or W/cm³). While the industry standard in recent years has been around 50W/in³ for high-efficiency PSUs with 80 PLUS Titanium certification, GaN-adopted designs are now seeing product examples exceeding 100W/in³. However, caution is needed when directly comparing these figures, as conditions vary with input voltage range, output voltage, and cooling method (air or liquid cooling).

The 48V power bus architecture, which is becoming prevalent as a characteristic power architecture for AI servers, allows for lower wiring currents than traditional 12V power distribution, thus reducing transmission losses. In this architecture, GaN FETs primarily handle the PFC (power factor correction) stage and DC-DC conversion stage, where 650V rated devices are well-suited. GaN adoption is increasing in 48V/3kW compatible PSUs that comply with the ORv3 standard promoted by the Open Compute Project.

This graph illustrates that GaN and Si compete in the 650V rated voltage range, and SiC occupies a segmented space for high-voltage applications above 1200V. The decision to adopt GaN in the PFC and LLC stages of AI server power supplies is made in comparison with Si MOSFETs of the same voltage class.

Cost Structure and Supply Realities

While the unit price of GaN FETs is still higher than that of Si MOSFETs, cases where GaN offers an advantage emerge when considering the overall system cost, which includes savings from reduced passive components and smaller PCB area. What is crucial here is not just the 'price of the individual device' but the 'cost viewed as the PSU's entire Bill of Materials (BOM).' A reduction in the number of passive components also alters implementation costs and reliability risks.

In terms of supply, while onsemi offers a portfolio of SiC MOSFETs, diodes, and modules covering voltages from 650V to 1700V, key suppliers in the GaN domain include EPC, GaN Systems (acquired by Infineon), Nexperia, and Texas Instruments. The supply chain for 650V GaN FETs for AI server power supplies has rapidly strengthened between 2023 and 2024. However, like SiC wafers, GaN also faces manufacturing yield challenges, making relationships with multiple suppliers a crucial factor for stable procurement in large volumes.

GaN device reliability evaluation may require tests beyond standard ALT (Accelerated Life Testing), such as High-Temperature Reverse Bias (HTRB) and monitoring for changes in dynamic RDS(on). Dynamic RDS(on) refers to the phenomenon where the on-resistance temporarily increases beyond its static value during switching operation, which is widely recognized as a challenge unique to GaN devices. The extent to which this phenomenon occurs depends on the device structure and manufacturing process, meaning that evaluating solely based on the static RDS(on) value in the datasheet may lead to an underestimation of actual operating losses.

So, What Specifically Should Be Confirmed?

With both technical and procurement aspects providing sufficient decision-making information, let's organize the points to confirm during the actual adoption consideration process.

The decision of whether to adopt GaN for AI server power supplies is not determined by a simple comparison of individual device characteristics. Rather, it requires a comprehensive assessment encompassing the entire power architecture, cooling design, and procurement risks. Just as short-circuit withstand time and DESAT design are core considerations for SiC selection, the suitability of dynamic RDS(on) and gate driver design becomes the central verification item for GaN selection. With the voltage range segmentation now becoming clear, a structural understanding of which material is advantageous for each system stage will expedite subsequent decisions on both design and procurement fronts.