Over the past few decades, metal oxide semiconductor field effect transistors (MOSFETs) have largely been the staple source for power supplies. MOSFETs are silicon-made devices controlled by voltage that manipulate the supply’s electricity that come in the form of little black squares. They’ve become very prevalent throughout the semiconductor industry, but might see their mainstream status begin to wither.
Emerging the scene is gallium nitride (GaN), devices that are expected to become smaller, cheaper, and more efficient in the long run. Silicon-based semiconductors had voltage coming into a data center at 48V go through multiple instances of power conversion before finally reaching its on-board components, during which the voltage would shed energy at each of these phases. According to Dr. Alex Lidow, chief executive of the Efficient Power Conversion (EPC) Corporation, Silicon wasn’t fast enough to reach 1V all the way from 48V.
“So what we (as an industry) did was create a whole bunch of very expensive power supplies that get you from 48V to 12V, and another set of power supplies that get you from 12V to 1V,” says Lidow. “And with gallium nitride, since it’s so damn fast, you can get rid of the whole intermediate bus and go directly from 48V to 1V.”
Although they’ve been under production for a few years, the initial costs of GaN chips were exorbitantly expensive. They were initially used on devices like LiDAR lasers and wireless power transmitters. Lidow’s company shifted gears and focused on data centers when the costs of manufacturing GaN chips fell. According to Lidow, gallium nitride has now become a “significant market” within the semiconductor industry, after years of convincing data centers on GaN’s true reliability and efficiency. Among data center operators (Facebook, Google, Microsoft, etc.), 48V power distribution is a major topic being thoroughly investigated. According to Lidow, many of these companies (without mentioning names) are conducting trial tests of GaN chips in their equipment.
To put the effectiveness of GaN into perspective, keeping a 580W machine powered actually requires the production of 860W, whereas the implementation of GaN chips would bring that figure down to 770W. Not only are GaN chips smaller than their silicon-based counterparts, but they produce less heat as well. Conventional distributive components of power take up significant space on server boards that would be better if designated to more CPU cores or RAM modules.
“What this means is a server has to communicate internally a whole lot more, and that puts a different metric on performance: there’s a big push to condense the boards and pack the servers much tighter, so that you can get this thinking process going,” says Lidow. “That is limited by the heat and power density of these systems, so you need more efficient power conversion—that’s one of your largest sources of heat.”
Filed Under: M2M (machine to machine)