Library

eCommons

 

Software, Architecture, and VLSI Co-Design for Fine-Grain Voltage and Frequency Scaling

Simple item page
dc.contributor.authorTorng, Christopher
dc.contributor.chairBatten, Christopher
dc.contributor.committeeMemberApsel, Alyssa B.
dc.contributor.committeeMemberManohar, Rajit
dc.date.accessioned2020-06-23T17:58:38Z
dc.date.available2020-06-23T17:58:38Z
dc.date.issued2019-12
dc.description156 pages
dc.description.abstractAs emerging domains demand higher performance under stringent constraints on power and energy, computer architects are increasingly relying on a combination of parallelization and specialization to improve both performance and energy efficiency. However, the combination of parallelism and specialization is also steadily increasing on-chip asymmetry in the form of spatial heterogeneity and temporal variation, which poses key challenges in the form of widely varying utilization in space (i.e., across different components) and in time (i.e., used at different times across varying performance levels). Fine-grain on-chip asymmetry requires analogously fine-grain power-control techniques in order to power (or not power) different components to different levels at just the right times to significantly reduce waste. At the same time, traditional walls of abstraction have broken down, allowing a cross-stack co-design approach across software, architecture, and VLSI to provide new, previously inaccessible information to precisely control new hardware mechanisms. This thesis explores novel fine-grain voltage and frequency scaling techniques to improve both performance and energy efficiency with software, architecture, and VLSI co-design. First, I explore architecture-circuit co-design and leverage recent work on fully integrated voltage regulation to enable realistic fine-grain voltage and frequency scaling for homogeneous systems of little cores at microsecond timescales. Second, I broaden the scope to heterogeneous systems of big and little cores and specialize for productive software task-based parallel runtimes. Third, I investigate much finer-grain asymmetry that can be exploited within coarse-grain reconfigurable arrays, which have recently attracted significant interest due to their flexibility and potential for reducing data movement energy. Finally, I describe my work on four silicon prototypes including a mixed-signal test chip and three digital ASIC test chips that support different aspects of my thesis. Throughout my thesis, I take a software, architecture, and VLSI co-design approach and focus on exploiting information newly exposed across layers of abstraction. I leverage a vertically integrated research methodology spanning across applications, runtimes, architecture, cycle-level modeling, RTL, VLSI CAD tools, SPICE-level modeling, and silicon prototyping to evaluate the potential benefit of fine-grain voltage and frequency scaling techniques.
dc.identifier.doihttps://doi.org/10.7298/97cd-fp69
dc.identifier.otherTorng_cornellgrad_0058F_11737
dc.identifier.otherhttp://dissertations.umi.com/cornellgrad:11737
dc.identifier.urihttps://hdl.handle.net/1813/69995
dc.language.isoen
dc.rightsAttribution 4.0 International
dc.rights.urihttps://creativecommons.org/licenses/by/4.0/
dc.subjectcomputer architecture
dc.subjectfine-grain dvfs
dc.subjectheterogeneous computing
dc.subjectintegrated voltage regulation
dc.subjectreconfigurable architectures
dc.subjectwork-stealing runtimes
dc.titleSoftware, Architecture, and VLSI Co-Design for Fine-Grain Voltage and Frequency Scaling
dc.typedissertation or thesis
dcterms.licensehttps://hdl.handle.net/1813/59810
thesis.degree.disciplineElectrical and Computer Engineering
thesis.degree.levelDoctor of Philosophy
thesis.degree.namePh. D., Electrical and Computer Engineering

Files

Original bundle
Now showing 1 - 1 of 1
Thumbnail Image
Name:
Torng_cornellgrad_0058F_11737.pdf
Size:
8.78 MB
Format:
Adobe Portable Document Format
Download

Collections

Cornell Theses and Dissertations