Design Power Converters with GaN Technology to Meet the Demands of AI Datacenters

– Barley Li, Applications Engineering Manager – Technical Content, APAC, DigiKey

The demand for artificial intelligence (AI) has rapidly increased datacenter energy requirements. To support increasingly complex AI workloads, leading server racks already consume more than 100 kilowatts (kW). This rise in server rack power density is making thermal management a critical design consideration.

Designers need semiconductor solutions that support high server rack power densities while reducing power conversion losses and associated heat generation.

This article explains why AI datacenter power requirements have grown and how this growth challenges designers. It then introduces gallium nitride (GaN) power solutions from onsemi and shows how they can be used to address these challenges.

Why AI has changed datacenter power requirements

Before the widespread adoption of AI, datacenters relied primarily on central processing units (CPUs) to handle a wide range of sequential workloads, including everyday web traffic and hosting, digital communications such as email, financial transactions, cloud storage and services, and enterprise database operations.

AI datacenters typically require a different approach. Graphics processing units (GPUs) provide a parallel processing architecture that supports the complex mathematical operations required to train and run large AI models quickly and efficiently. However, this performance can come with higher peak power consumption.

CPUs still play a distinct and vital role in AI datacenters, handling data preprocessing, assigning AI workloads across GPU clusters, moving data, and managing I/O operations. To keep pace with GPUs and prevent data bottlenecks, CPUs must deliver sufficient processing performance, memory bandwidth, and I/O throughput. In some cases, this can further increase AI server rack power consumption. 

Furthermore, higher-bandwidth network interfaces are required to meet the substantial demands of AI processing, especially when moving data between server racks. This can increase power consumption in several ways, depending on interface power usage, link rates, and whether network transceivers use copper or optical connections, which present power and thermal tradeoffs based on transmission distance.

The processing and networking requirements of AI server racks have led many engineers to rethink datacenter power design, a task further complicated by the close link between processor power consumption and heat generation.

Power design challenges in AI server racks

AI datacenters must deliver power to a wide range of heterogeneous components. To enable more flexible power design, engineers are beginning to disaggregate three-phase AC-DC conversion from individual processing racks, instead housing this infrastructure in dedicated sidecar racks. This allows datacenter engineers to partially decouple processing resources from power infrastructure.

Depending on the architecture, AI datacenter power conversion can involve several key stages:

• Front-end conversion takes three-phase AC and typically uses power factor correction (PFC), rectification, and electromagnetic interference (EMI) filtering to produce a high-voltage DC bus that serves as a baseline for subsequent downstream conversion.
• High-frequency isolated DC-DC conversion steps these voltages down to the levels required by downstream AI resources. Distributing power and providing high-current, low-voltage conversion closer to the load can reduce resistive distribution losses and improve energy efficiency. Galvanic isolation electrically separates the converter’s input and output, helping meet safety and system protection requirements.
• Intermediate bus conversion (IBC) takes a rack-level DC supply and converts it to an intermediate voltage suitable for distribution within the server rack. For example, a 48 volt nominal distribution bus, which may operate at approximately 54 volts, can use IBC to produce a 12 volt rail before the final conversion to processor voltages.

• Point-of-load (POL) converters deliver the final operating voltage to individual processors. They are placed near the target components to reduce electrical noise, distribution losses, and parasitic impedance, while supporting fast transient response when power demands change rapidly.

Together, these conversion stages condition and regulate power for each processing resource across the server racks. However, because conversion losses generate heat, designers benefit from power semiconductors that enable efficient voltage conversion and reduce the thermal management burden.

Yet, designers face another challenge: rising performance expectations for a given footprint of datacenter equipment. Removing front-end conversion from server racks can free up space for additional processing resources, boosting overall processing density. Still, these resources require later-stage power conversion that must fit within limited space.

By choosing suitable power semiconductor technology, designers can minimize losses and reduce the power converter footprint, enabling higher processing and power density within each server rack.  

Increasing AI server rack power density and conversion efficiency with GaN

GaN is a wide-bandgap (WBG) semiconductor material that offers many advantages for power applications, such as voltage conversion. Its high-frequency switching performance allows designers to use smaller filtering components, including magnetics that typically occupy significant printed circuit board (pc board) area. By minimizing the footprint of power converter circuitry, designers can increase power density and offer greater processing resources per AI server rack.

GaN also exhibits low switching losses, which can improve conversion efficiency and reduce heat generation. While this does not reduce the heat generated by the processors themselves, GaN-based power conversion helps reduce the additional thermal management burden imposed by AI server power infrastructure.

To address the various stages of power conversion across AI datacenters, designers benefit from a comprehensive GaN power portfolio. The onsemi GaNEXUS power portfolio currently features lateral GaN devices covering 40 to 650 volts, including discrete enhancement-mode GaN high-electron-mobility transistors (HEMTs) and devices with integrated protection features.

Because AI datacenters represent substantial long-term investments, AI server rack designers need solutions that ensure high reliability and long lifecycles. onsemi designed the GaNEXUS portfolio around these requirements, applying WBG semiconductor expertise.

Exploring the GaNEXUS product portfolio

By selecting specific components from the GaNEXUS portfolio, designers can address the key stages of power conversion in AI datacenters. For front-end AC-DC and high-frequency isolated DC-DC conversion, the ENGNTBT035N65GN1TXG (Figure 1) is a 650 volt enhancement-mode GaN HEMT in a 10 × 15 millimeter (mm) surface-mount TOLT package for top-side cooling in high-density designs.

This device features a low typical ON resistance (R_DS(ON)) of 27 milliohms (m?) to minimize conduction losses during power conversion, while supporting a maximum continuous drain-source current (I_DS) of 64 amperes (A) at a case temperature (T_CASE) of 25°C.

The ENGNTBT035N65GN1TXG’s ultra-low gate charge (Q_G) of 14 nanocoulombs (nC) enables high switching speeds, supporting higher power density with smaller magnetics.

Additionally, the ENGNTBT035N65GN1TXG can withstand non-repetitive drain-to-source transient voltages (V_DS(TRAN)) of 800 volts, providing additional voltage headroom during short-duration abnormal power events.

Alternatively, the ENGNCP58934ABLTXG (Figure 2) is also a 650 volt enhancement-mode GaN power switch with similar current capabilities and R_DS(ON), but it comes in a 10.38 × 9.90 mm PDSO-F9 package with maximum I_DS of 63 A at a case temperature of 25°C. This device integrates a gate-voltage clamp and a Miller clamp and supports adjustable turn-on and turn-off slew rates via external resistors.

These features help designers manage GaN gate robustness and balance EMI performance with conversion efficiency. Note that an external gate driver is required.

For intermediate bus conversion, the ENGNTLCC3D5N10GN1TXG (Figure 3) is a 100 volt device that supports a maximum continuous I_DS of 183 A at a case temperature of 25°C. It offers a low R_DS(ON) of 2.7 m? and an ultra-low Q_G of 7.3 nC, reducing losses in later-stage power conversion.

The ENGNTLCC3D5N10GN1TXG is available in a 3.3 × 3.3 mm PTFP-N9 package, highlighting the power-density potential of GaN-based power converters.

Further downstream, the ENGNTLEF0D7N04GN1TXG (Figure 4) is a 40 volt device with a maximum I_DS of 819 A at a case temperature of 25°C. It is suitable for high-current DC-DC conversion near the load. It comes in a 5 × 6 mm PDSO-N8 package, and its typical R_DS(ON) of 0.5 m? helps minimize losses during high-current operation, supporting efficient power conversion and reducing heat generation.

Conclusion

AI datacenters typically require multiple stages of power conversion within increasingly complex server racks. By using GaN power semiconductors in high-frequency front-end and DC-DC converters, designers can reduce conversion losses and the associated heat while shrinking the converter footprint.

The GaNEXUS power portfolio from onsemi offers high-performance GaN solutions that meet electrical, packaging, thermal, and integrated protection requirements across multiple stages of AI datacenter power conversion.