AI's Edge: Fueled by Power Electronics

With its ability to substantially increase business efficiency, AI is poised — and has already begun — to have the same magnitude of impact as the arrival of the internet in the late 1990s. However, since something as simple as a generative AI query will use almost ten times as much electricity as a search engine query, the rapid growth of AI use is expected to drive a 160% increase in data centre power demand by 2030. By some estimates, data centres already consume up to 3% of the world’s electricity, so there is enormous pressure on operators to reduce energy consumption and increase use of renewables.

The rising tide of AI

Power conversion is a core component at all levels of the Cloud-Edge architecture. Large data centres are traditionally powered by 3-phase AC voltages, converted to stable DC voltages for distribution to the various racks and servers within the centre.

This model is changing as data centre operators integrate alternative energy sources into their power configurations and so DC-to-DC and AC to DC conversion techniques are employed to integrate solar, wind, and battery storage into grid supplies.

At the same time, the twin challenges of packing more computing power into the same rack space, while decreasing energy consumption, are throwing the spotlight on the efficiency and density of power supplies. The numbers of individual power supplies within the data centre are being reduced as DC power is distributed directly to the server racks, and a trend towards distribution at higher voltage is reducing losses. Efficient DC-DC conversion is essential, therefore, and converters must be able to withstand the higher distribution voltages.

At the edge, the demands on power supply design are equally challenging as AI is leveraged by an ever-increasing variety of applications, ranging from electric vehicles to industrial automation. Power solutions at the edge must therefore cope with a wide variety of operating environments, often hostile, and voltage sources, including battery power.

AI demands innovation in electronic power design

Engineers implement power conversion using a range of topologies based on switching and passive components, with multiple topologies often combined to improve functionality and reliability. The silicon MOSFET has traditionally been the device of choice for these topologies, but the performance and efficiency requirements of next-generation power supplies for AI applications are exceeding the capabilities of silicon, and designers are turning to wide bandgap materials, such as silicon carbide (SiC) and gallium nitride (GaN).

Both materials enable the fabrication of power semiconductors, which are smaller and lighter than silicon devices and enable higher switching frequencies, improving conversion efficiency and shrinking the sizes of passive components — and hence overall form factors. Additionally, since higher power densities and conversion efficiencies lead to a reduced need for thermal management, using wide bandgap technologies further eases the pressure on physical real estate.

Both SiC and GaN also have higher withstand voltages, SiC as much as 1200V, while GaN is typically 600V. GaN is therefore particularly suited to data centre and industrial power supply design, offering significant opportunities to reduce power consumption. GaN devices, however, have relatively fragile gates and careful gate driver design is required to avoid problematic high and low-side ringing, potentially damaging shoot through currents.

Designers are therefore faced with several decisions when developing solutions for AI infrastructure. An ever-increasing range of GaN and SiC devices must be evaluated against the requirements of the application and the designer must select the best topology for the application. Gate driver design is crucial for both SiC and GaN devices as it significantly influences the safe operation of the device, and some GaN devices now come with integrated gate drivers to simplify the design challenge.

With pressure on development cycles increasing, the developer does not have the luxury of time when weighing the above factors and must rapidly make the best decisions for the optimal solution design. But poor decisions in the design phase can have severe consequences later in the design cycle — so, how can you be confident that your design choices are fully informed?

Featured AI Solutions