As deep learning proliferates, the question of data center power density is once again on the rise, creating new business opportunities for specialized cloud services, hosted in facilities that can support north of 30 kW per rack, and companies in the power conversion space, who can tackle the density issue by making systems more energy efficient.
A material that’s enabling better vision in self-driving cars, a better augmented-reality experience, and wireless power, is also promising extreme energy efficiency improvements in the data center. Replacing silicon as the semiconductor material in power conversion chips with gallium nitrate, or GaN, leads to much smaller and more energy efficient devices that provide much faster switching.
Training deep neural networks today requires powerful computers filled with GPUs. These machines need a lot of power, and data centers that host them have to be designed for high power density, similar to the way supercomputer data centers are designed. The radical energy efficiency improvements promised by GaN power converters on motherboards mean more computing power can be stuffed in a single data center cabinet, making the technology appealing to a whole range of companies in the data center space, from those in the hardware supply chain to operators of hyper-scale cloud platforms.
Cutting Conversion Losses in Half
Interest in what GaN devices can do for data center power density is “high across the board,” Steven Tom, product line manager at Texas Instruments, said in an interview with Data Center Knowledge. According to him, the material “halves power losses” when used instead of silicon.
A company that has been a major force behind GaN chips in the power conversion market is El Segundo, California-based Efficient Power Conversion. One of EPC’s two founders is Alex Lidow, who is probably GaN’s most prominent and enthusiastic evangelist. In the semiconductor business for 40 years, he was one of the inventors behind the power MOSFET, a silicon transistor widely used for power switching in servers and many other types of electronics. Now he’s advocating for GaN transistors to push those silicon MOSFETs out.
TI, one of the world’s largest semiconductor companies, partners with EPC for its GaN devices, which it uses to build System-on-Chips, or SoCs. An SoC is essentially a single device that’s made up of multiple devices of different types. TI’s upcoming product for the data center market, for example, combines two EPC GaN transistors and other components into a single power-conversion chip.
Hyper-Scale Cloud Firms Dabble in GaN
TI has seen requirements come in for power distribution solutions for motherboards with multiple GPUs, Tom said. When they come in, customers don’t usually specify that they want GaN to be used, but the material has proven to be effective in addressing those requirements.
Operators of web-scale platforms hosted in massive data centers – companies that according to Lidow include Facebook, Google, and Oracle – have been buying GaN chips and exploring the technology, although they are not yet building it into their servers at scale. Artificial Intelligence and cloud are making data center power density an acute issue for these firms, and GaN-based power conversion is one potential solution, he said.
Silicon-based server power conversion device (left) vs. a GaN-based one (right). The blue chips on the smaller device are EPC's GaN transistors (Photo: Yevgeniy Sverdlik)
LMG 5200, the aforementioned power controller for servers with EPC’s chips at the heart, is TI’s first GaN device for the data center market. EPC chips are in “several major servers” expected to ship this year, a company spokesperson wrote in an email. The volume of devices shipped for server use up to this point has been “less than a million units.”
The second EPC founder is Archie Hwang, owner of Episil Technologies, which is based in Taiwan and has a silicon foundry there. EPC builds its wafers in that foundry and grows GaN crystals in a facility on the premises. That’s one of the ways the company has been able to bring down the cost, and it’s possible because GaN is grown on top of silicon – a process invented in 1999 by the Japanese scientist Hiroyasu Ishikawa. EPC also covers the layer of GaN transistors in its devices with multiple layers of glass, eliminating the need for processor packages, which constitute more than half the cost of silicon chips, according to Lidow. Finally, the material’s much higher density enables the company to fit many more transistors on a single wafer than silicon would allow.
EPC and TI aren’t the only GaN game in town. Their competitors include Transphorm, a Goleta, California-based maker of GaN power conversion chips that has attracted hundreds of millions of dollars in funding from investors such as Google Ventures and Kleiner Perkins Caufield and Byers, among others (Lidow and Hwang are EPC’s two sole financial backers). Another example is Macom, a Lowell, Massachusetts-based semiconductor company that three years ago acquired Nitronex, one of the first companies to productize GaN on silicon.
Lidow and Tom are both extremely enthusiastic about the future of GaN in the broader power conversion industry. “We think it’s a complete game-changer,” Tom said, describing TI’s position. “We have a very strong opinion of the technology; it really does things silicon as a technology just can’t reach.”
People in the semiconductor industry have known for a long time that between GaN and silicon, GaN is the better semiconductor material. But until Ishikawa’s breakthrough, which made it possible to produce GaN wafers in the same fabs that produce silicon wafers, the cost of GaN chips made them infeasible, Lidow said.
It’s a better semiconductor material because of its density. Because the chemical bond between atoms in GaN is stronger than in silicon, things can be brought much closer together without losing control of the flow of the electrons.
Compared to silicon, GaN allows for fewer unwanted imperfections in the electrical waveform inside the power supply. “In a power supply it’s all about making perfect square waves,” Tom said. What prevents those perfect square waves from forming is dissipation: nasty rings, overshoots, undershoots, all that energy circulates on the board. With GaN chips “we can make textbook waveforms.”
The improvements don’t come simply as a result of replacing silicon transistors with GaN ones. The devices are completely redesigned from the ground up. TI’s new devices cut the amount of conversion steps energy goes through between the point of entry and its destination. Steps like AC input, isolation, intermediate bus converter, and point-of-load converter (where each step is accompanied by some power loss) are reduced to two: from 48V DC to 12V DC and from 12V to 1V at each point of load. “You get the same performance or better in that conversion because of the device properties,” Tom explained.
From Beetle to Ferrari
Lidow has no doubt GaN will disrupt a large portion of the $30 billion power conversion industry. “As sure as the sun comes up, GaN is going to replace silicon in power conversion,” he said. It is, however, unlikely to replace silicon in logic semiconductors (CPUs and GPUs). “It’s possible, but it’s not certain.”
The only disadvantage of GaN he sees today is the lack of knowledge about it. “Where it loses to silicon is in the user base.” Nothing has ever come along that was so superior to silicon in cost and performance, and switching from one to the other is like switching from an old VW Beetle to a Ferrari, Lidow said, a switch that comes with a learning curve for the driver.