This increase isn't solely due to the greater number of data centers; their power density is also rapidly increasing. The power consumption of the most advanced GPUs (graphics processor units) used for AI is above 1 kW, a significant increase from the 500 W of the previous generation. The higher power demands of these components, coupled with the pressure to increase the number of GPUs per rack, are pushing the power consumption of a single rack to as high as one megawatt.
The power density of these types of racks is evident when you consider that, although they are the size of a large refrigerator, it is expected that by 2027 the consumption of a single rack will be equivalent to that of 65 homes.
This energy demand has driven the evolution of the electrical distribution network and the necessary power supplies. Instead of relying on an AC power distribution system installed in the data center via a 48V power bus, Nvidia now proposes a DC power supply distributed to the AI servers in the data center at 800V. This requires designing new, smaller, and more efficient power supplies to reduce losses.
These trends have a huge impact on power distribution and conversion in the data center. The need to distribute gigawatts of power necessitates reducing losses and increasing the efficiency of power supplies. Increasing efficiency by just a couple of percentage points, from 97% to 99%, can significantly reduce power consumption and offer operational cost savings and environmental benefits. Higher efficiency also means less cooling is required, saving on water and dielectric materials, and allowing for smaller power supply systems. New packaging technology can also simplify cooling system design, further reducing the size of the power supply and enabling more GPUs to be installed in a rack.
Gallium nitride (GaN) technology offers the possibility of achieving higher efficiency in a smaller area thanks to its ability to switch at higher frequencies than other semiconductor technologies.
However, adopting this technology is not as simple as directly replacing silicon devices in power supplies. GaN requires more complex gate control, which varies between -3 V and +6 V. This means that different device drivers with more complex timing are needed.
ICeGaN
Adding a gate interface to a 650 GaN transistor with low on-resistance (RDS(on)) can greatly simplify the device's use in AI server power supply design. This allows a standard MOS driver chip with a typical bias of 12 V to 15 V to be used to drive the GaN.
The ICeGaN® technology developed by Cambridge GaN Devices (CGD) consists of a power integrated circuit that incorporates this interface on the same substrate as the enhancement-mode HEMT transistor, instead of using a separate silicon MOSFET in a cascode configuration or a silicon integrated circuit within the same package. Implementing the interface on the same chip allows for an optimized, low-profile package with low inductance, a fundamental requirement for reducing losses and increasing the efficiency of the power converter.
ICeGaN devices incorporate an auxiliary gate HEMT that is attached to the external gate input and provides a stable internal gate bias to the main HEMT in enhancement mode. This means that ICeGaN devices are controlled with gate biases between approximately 9 V and 20 V and can safely withstand higher transient voltages. They also double the apparent threshold voltage (VTH) to about 3–4 V, thus improving gate noise immunity.
A Miller clamp circuit is also integrated into ICeGaN devices to provide a strong buck boost when the HEMT is not conducting. This eliminates the need for designers to supply negative gate voltages to keep the device off when it is supposed to be in this mode. The Miller clamp circuit also widens the safe operating area (SOA) and increases gate reliability by absorbing steeper dV/dt slopes. This allows for increased switching speed without the need for additional negative voltage rails. By relying on gate resistance variation during on-conductivity control, designers can control the ICeGaN switching speed according to application requirements, achieving a balance between EMI and switching losses.
Dynamic GPU power consumption for AI workloads causes large, rapid load fluctuations that cannot be fed back into the AC grid. To prevent this, each stage of the AC/DC conversion chain must be able to handle transient loads and include sufficient local energy storage to smooth out the corresponding power fluctuations.
This involves sizing components and thermal design based on peak loads, not average loads, to ensure that the 800V DC/DC converter can supply higher voltages for short periods without suffering excessive stress or loss of regulation.
The voltage at the inner gate of the power HEMT begins to increase when the voltage at the outer gate exceeds approximately 1.5 V. This additional threshold is what allows ICeGaN to offer a higher threshold voltage than conventional GaN in discrete enhancement mode at 2.9 V.
One of the main advantages of GaN HEMT devices is their switching performance with a very low gate load and no structural diode, resulting in virtually no losses when switching to cutoff. Because the Miller blocker turns the device off at 0 V, ICeGaN devices do not suffer from the high reverse conduction losses common with other GaN components. Even more importantly, device cooling can be significantly improved by grounding the device, allowing for smaller components and lower operating temperatures.
Encapsulation
Reducing the thermal resistance of the package offers several advantages. More power is available for the same RDS(on), and devices can operate at lower temperatures for the same power output, allowing for a smaller, lighter heatsink and thus lowering system costs. Lower operating temperatures also increase reliability, extend lifespan, and reduce the size of power supplies in AI racks.
The BHDFN-9-1 (bottom heat-spreader DFN) package developed by CGD is bottom-cooled and features wettable edges for easy inspection. It utilizes a thicker leadframe for heat dissipation, providing a thermal resistance of 0.28 K/W, a value that matches or exceeds that of other devices. Measuring 10 x 10 mm, the BHDFN package is smaller than the commonly used TOLL package but occupies a similar footprint, making it easy to use.
Unlike discrete GaN devices, ICeGaN has an additional pin for powering its internal circuitry, designated pin 1, while the gate terminal is pin 2 and the Kelvin source (KS) terminal is pin 3. In the TOLL package, the gate terminal is pin 1 and the KS terminal is pin 2. This facilitates the development of common and compatible layouts, in which case a TOLL device is moved aside to avoid connecting to the power terminal in the ICeGaN layout. This allows for almost direct replacement of a TOLL device with an ICeGaN device.
Design of 800V power conversion systems with GaN devices
A commonly used topology for powering data centers consists of a stacked half-bridge resonant converter in an ISOP (input-series, output-parallel) configuration that allows the conversion of 800 VDC to the input using 650 V GaN devices. Flat transformers and advanced magnetic integration are essential to provide the compact, low-profile power stages needed when the converter modules must be located close to the GPU and CPU boards.
800 V distributed power networks are already using switching frequencies in the 500–800 kHz range with reference designs for AI server applications. Alongside the more widespread adoption of GaN technology, there is a strong push towards 1 MHz and beyond to further increase power density and reduce the size of magnetic components, provided that efficiency and EMI levels can be maintained at these higher frequencies.
Security
Safety is a fundamental aspect for 800 VDC DC/DC architectures and this requires taking other factors into account for their design.
At 800 VDC, it is essential to reinforce the insulation, increasing the air clearances and creepage distances compared to bus architectures with lower voltages. ±400 VDC systems use a high-resistance or solid ground connection at the midpoint to the protective earth and require a shorter creepage distance in the insulation than 800 V systems with simple termination.
The hot swap and eFuse functions are usually integrated with the 800 VDC stage on a distributed power board so that isolation from current spikes, overloads, and faults is controlled as close as possible to the power supply of each rack.
800 V DC/DC converters are also often implemented as open-loop LLC stages with the controller on the secondary side and limited sensing on the primary side. This means that protection on the primary side, such as transient current and abnormal state detection, is required to safely handle input faults.
In the event of a failure, the design of the 800VDC DC/DC converter must strictly contain the power to prevent smoke and fire from spreading to GPU boards or other components, thus protecting both system uptime and hardware integrity.
A real-world example
A prototype 3 kW, 400 V to 50 V half-bridge module has been built using CGD P2 devices with a 25 mΩ on-resistance on the primary side of a planar transformer to demonstrate the high levels of power density and efficiency achievable with HEMT ICeGaN. The module achieves a peak efficiency of 98.7% and delivers 97.5% efficiency at full load.
The modular architecture facilitates parallel configuration for higher power levels, such as 6 kW or 12 kW, in high-power DC/DC applications. Phase-shedding between interleaved modules optimizes overall system efficiency across a wide range of loads.
Conclusion
Power supply designers are increasingly considering GaN technology for data center applications. However, achieving optimal performance can be challenging when using conventional GaN devices. For example, the required negative gate voltage limits operation at the higher frequencies achievable with smaller power converter-based designs.
However, incorporating an intelligent interface on the same chip as the GaN power device brings several advantages to AI data center power designers, as it greatly reduces the cost, size, and weight of power supplies in 800V data center power networks.
Author: Di Chen, Technical Marketing Director, Cambridge GaN Device https://www.camgandevices.com
