Architecture level power-performance trade-offs in data-dominated designs

Abstract

As the demand for feature-rich portable devices continues to increase, new techniques are needed to minimise power consumption. This thesis is concerned with the development and validation of new systematic architectural methods of determining pipeline stage insertion in data-dominated designs with the aim of reducing dynamic power consumption. The methods place special emphasis on the number of latches used in pipeline stages and voltage scaling. The first part of the thesis addresses power minimisation through systematic analysis of the number of latches in pipeline stages. A new pipeline stage insertion (PSI) method operating at the architectural level is developed which takes into account system clock period and FEs outputs and delays. A PSI algorithm based on analytical heuristic equations is formulated to ensure the successful application of this method to any given data-dominated design. The input to the algorithm is designer clock period and naively inserted pipeline stages. The output from the algorithm is a pipelined design fulfilling the timing constraint with the least dynamic power consumption. To support efficient power-performance trade-offs exploration, the algorithm was fully automated. The second part of the thesis focuses on the validation of the PSI method using two real-life case studies: triple data path floating-point adder and MPEG-1 motion compensation module. These designs are common in many portable devices and have numerous implementation challenges large number of FEs and significant power consumption. Extensive experimental results show that for the motion compensation module, the PSI is able to reduce dynamic power consumption by up-to 30% compared with other reported approaches. The final part of the thesis concentrates on voltage scaling (VS) and its impact on pipeline stages. The timing slack available in each stage is investigated, with the aim of further reducing power consumption by lowering the supply voltage. The PSI method is modified to support voltage scaling, and as a result, a new pipeline stage insertion with voltage scaling (PSI-VS) method is proposed. Experimental results show that the PSI-VS can lead to significant power saving compared with PSI without VS. For the MPEG-1 motion compensation case study, a power saving of 68% is observed. All the developed methods have linear time complexity as the number of pipeline stages increases, facilitating their application to large designs without incurring run time penalty. The results for the case studies were based on a synthesisable RTL implementation using 90nm technology together with accurate power analysis using commercial tools.

More information

Identifiers

Catalogue record

Export record

ASCII CitationAtomBibTeXData Cite XMLDublin CoreDublin CoreEP3 XMLEndNoteHTML CitationHTML CitationHTML ListJSONMETSMODSMPEG-21 DIDLOpenURL ContextObjectOpenURL ContextObject in SpanRDF+N-TriplesRDF+N3RDF+XMLRIOXX2 XMLReferReference ManagerSimple Metadata