Abstract
dc:descriptionComputing systems in almost every application domain now support techniques to trade off power and performance. Such techniques are used to enforce power and thermal constraints, manage power and thermal budgets and respond to temperature and aging. Unfortunately, many of the current techniques are limited in the dynamic range they provide and scale poorly with technology. Techniques that can supplement or replace current techniques are needed. We propose k-hot pipelining, a novel technique to support multiple power-performance points in a processor. The key idea is to provide power and clock to only k stages of an m-stage pipeline (k < m); the k stages to be powered on change as instructions flow through the pipeline. Since the remaining m − k stages do not consume power, the technique results in power savings at the expense of performance. k-hot pipelining can be software or hardware-controlled, workload-agnostic or workload-adaptive, and can be used to provide power-performance points not supported by existing techniques. For one implementation of k-hot pipelining, we show that up to 49.9% power reduction is possible over the baseline design. Power reduction is up to 47% over the lowest power point supported by DVFS.
Degree
thesis:*- Name thesis:degree_name
- M.S.
- Level thesis:degree_level
- Thesis
- Discipline thesis:degree_discipline
- Electrical & Computer Engr
- Grantor
- University of Illinois at Urbana-Champaign
- Year dc:date
- 2016
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Lu, Dao
- Contributors dc:contributor
-
- Kumar, Rakesh
Subjects
dc:subject × 2Rights
dc:rights- Statement dc:rights
-
- Copyright 2016 Dao Lu
- Language dc:language
- en
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/2142/93032
- OAI identifier oai:identifier
- oai:www.ideals.illinois.edu:2142/93032