Optimality Study ofResource Binding withMulti-Vdds

Deming Chen, Yiping Fan, Junjuan Xu

Abstract

Deploying multiple supply voltages isoneofthemosteffective Deploying multiple supply voltages (multi-Vdds) ononechip isan techniques toreduce dynamic powerunder performance constraint. important technique toreduce dynamic powerconsumption. Inthis Thistechnique hastheadvantage ofreducing powerdissipation workwepresent anoptimality study forresource binding targeting without sacrificing theperformance ofthesystem. Whenthe designs with multi-Vdds. This issimilar tothevoltage-island design threshold voltage forthetransistors stays asaconstant, thedelays of concept, except that thegranularity ofourvoltage island isonthe theresources becomelonger asthesupply voltage scales down. functional-unit level asopposed tothecorelevel. Weareinterested Therefore, wecanassign high-Vdd tocritical paths tomaintain high inachieving themaximumnumber oflow-Vdd operations and, in performance andlow-Vdd tonon-critical paths tosavepower. thesametime, minimizing switching activity during functional uiit Clusters ofhigh-Vdd cells andlow-Vdd cells werefirst usedin[21]. binding. Tothebest ofourknowledge, there isnoknownoptimalTheworkin[20] adopted multiple supply voltages inthereal design solution tothis problem. Tocompute anoptimal solution forthis ofaMPEG4video codec. Researchers areworking from various problem andexamine thequality gapbetween oursolution and abstraction levels toexplore theadvantages ofmultiple supply previous heuristic solutions, we formulate thisproblem asa voltages. Specifically, there areworks onsystem level [10][12][18], min-cost network flowproblem, butwithspecial equal-flow behavioral level[7][11][14][16], logic level[8][9], andphysical constraints. Thisformulation leads toaneasyreduction tothe ' .'. integer linear programming (ILP) solution andalso enables efficient level.Ou work fceonpoeoptimization at the approximate solution byLagrangian relaxation. Experimental behaircal thedesigngeneral,vthe igr thedesign llthe mor e results showthat theoptimal solution computed based onour crtcalthe desgndecsions arehfor thequality ofthe final product. formulation provides 7%morelow-Vdd operations andalso reduces Theessence ofbehavioral synthesis withmultiple supply thetotal switching activity by20%compared tooneofthebest voltages istoassign low-Vdd values toasmanyoperations as knownheuristic algorithms thatconsider multi-Vdd assignmentspossible under latency andresource constraints. In[17], anoptimal only. solution wasgiven fortime-constrained scheduling problem under variable voltages. However, itdidnotconsider resourceconstraints.
BibTeX
@inproceedings{chen2006optimality,
  author = {Deming Chen and Yiping Fan and Junjuan Xu},
  title = {{Optimality Study ofResource Binding withMulti-Vdds}},
  booktitle = {DAC},
  year = {2006},
  doi = {10.1145/1146909.1147059},
}

Loading related content...

Figure