Wide Bandgap Conversion at the Shelf
Gallium nitride at five hundred kilohertz, and what the frequency buys
Why switching faster makes a power shelf smaller. Where the gain is spent rather than banked. What a 45 degree inlet does to the whole argument.
Abstract
A power shelf converts 800 volts direct current into what servers actually eat. This paper sets out why HyperNext chose gallium nitride running near five hundred kilohertz for that conversion, what the higher frequency does to the magnetic components that dominate the volume of any converter, where the efficiency gain is spent rather than kept, and why designing for a 45 degree inlet rather than a 25 degree one changes the device selection before anything else is decided.
Contents
- 011. What a power shelf has to do
- 022. Following Open Compute, and where we go past it
- 033. Silicon, silicon carbide, gallium nitride
- 044. The five hundred kilohertz decision
- 055. Where the frequency gain is spent
- 066. Designing for a forty-five degree inlet
- 077. Making it in India
- 088. References and standards
Request the full paper
The complete paper, including all figures, tables, references, and citations, is available as PDF. Enter your details to receive it.
Request paper · HN-RP-013.pdfKey findings
- Transformer and inductor volume falls roughly with frequency for a given power, which is the whole reason to switch faster. Nothing else in the converter shrinks nearly as much.
- Gallium nitride has essentially no reverse recovery charge, which removes the loss mechanism that made high frequency uneconomic in silicon.
- The DC-DC shelf reaches about 97.5 per cent at full load with a higher peak at partial load. The DC-AC shelf reaches about 98.3 per cent measured at the server.
- Designing for a 45 degree inlet rather than a comfortable one costs junction temperature margin, and that cost is paid in device selection and board area rather than recovered later in the thermal design.
- Higher frequency is not free. It moves loss out of the magnetics and into gate drive, layout parasitics and electromagnetic compatibility, and each of those has to be engineered rather than assumed away.
