Applying Differential Power Processing to Series-Stacked Processors for Efficiency Improvement in a Multi-Processor Server or System
PhD student Andrew Stillwell with advisers R. Pilawa-Podgurski and P. Sauer
As the capabilities of the internet and cloud services grow, so too grows the demand for data centers. These centers are increasingly taking up a larger percentage of the US energy capacity, leading to a need for efficient power delivery to servers. One growing trend in servers is the move to many, smaller, lower power processors in a compact form factor with the intent to lower power consumption and decrease footprint. Prior research has focused on servers electrically connected in series to increase data center power efficiency by removing a power processing step and applying differential power processing (DPP) to balance the servers. This research investigates the application of DPP to series-stacked processors to improve the efficiency of a multi-processor server or system.
Differential power processing allows for highly efficient power conversion by processing only the difference in load currents of series-connected loads. Switch-capacitor (SC) converters are switching power converters that utilize capacitors rather than inductors as their primary energy storage element. A prototype SC converter to balance a stack of four ARM embedded computers was developed as shown in Figure 1. Over the last few months of testing, the converter reached a peak conversion efficiency of 98.1% and a peak power delivery efficiency of 99.8%. In addition, a light load operation mode was implemented that reduces power conversion losses by reducing the switching frequency. Figure 2 shows efficiency improvement using light-load operation for well-matched loads. This work is funded by the UIUC College of Engineering, Strategic Research Initiative.