Computer Power Supply Transient Response Improvement by Power Consumption Prediction Procedure Using Performance Counters

Shinichi KAWAGUCHI  Toshiaki YACHI  

Publication
IEICE TRANSACTIONS on Communications   Vol.E98-B   No.12   pp.2382-2388
Publication Date: 2015/12/01
Online ISSN: 1745-1345
DOI: 10.1587/transcom.E98.B.2382
Type of Manuscript: PAPER
Category: Energy in Electronics Communications
Keyword: 
performance counter,  power supply,  transient response,  power consumption,  POL,  regression analysis,  

Full Text: PDF>>
Buy this Article




Summary: 
As the use of information technology is rapidly expanding, the power consumption of IT equipment is becoming an important social issue. As such, the power supply of IT equipment must provide various power saving measures through advanced features. A digitally controlled power supply is attractive for satisfying this requirement due to its flexibility and advanced management functionality. However, a digitally controlled power supply has issues with its transient response performance because the conversion time of the analog-digital converter and the time required for digital processing in the digital controller adversely affect the dynamic characteristics. The present paper introduces a new approach that can improve the transient response performance of the digital point-of-load (POL) power supplies of computer processors. The resulting power systems use feed-forward transient control, in addition to the general voltage regulation feedback control loop, to improve their dynamic characteristics. On the feed-forward control path, the processor workload information is supplied to the power supply controller from the processor. The power supply controller uses the workload information to predict the power load change and generates an auxiliary control to improve the transient response performance. As the processor workload information, the processor-integrated performance counter values are sent to the power supply controller via a hardware interface. The processor power consumption prediction equation is modeled using the moving average model, which uses performance counter values of several past steps. The prediction equation parameters are defined by multiple regression analysis using the measured CPU power consumption data and experimentally obtained performance counter information. The analysis reveals that the optimum parameters change with time during transient periods. The modeled equation well explains the processor power load change. The measured CPU power consumption profile is confirmed to be accurately replicated by the prediction for a period of 200ns. Using the power load change prediction model, circuit simulations of the feed-forward transient control are conducted. It is validated that the proposed approach improves power supply transient response under some practical server workloads.