The gap created by diminishing benefits from technology scaling and projected growth in computing demand has led to a need for new sources of computing efficiency. Fortunately, the workloads that are driving demand across the computing spectrum also present new opportunities. In data centers and the cloud, computing demand is driven by the need to organize, analyze, interpret, and search through exploding amounts of digital data. In mobile and deeply embedded devices, the creation and consumption of richer media and the need to interact more naturally and intelligently with users and the environment drive much of the computing demand. These applications are largely not about calculating a precise numerical end result; for them, "correctness'' is defined as producing results that are good enough, or of sufficient quality. How does this help design more efficient computing platforms? These emerging workloads, and many others, demonstrate a high degree of intrinsic resilience to their underlying computations being executed in an approximate or inexact manner. This project will explore approximate computing, an emerging design approach that improves the efficiency of computing platforms by designing them to leverage intrinsic application resilience.

To establish approximate computing in a broader context, this research explores quality programmable processors - programmable platforms for approximate computing that offer the ability for software to express application resilience at the natural HW/SW interface, i.e., the instruction set. The hardware underlying a quality programmable processor is equipped to understand and guarantee the instruction-level quality specifications, while exploiting the flexibility that they provide to obtain performance or energy improvements. This project will explore quality programmable designs of various programmable architectures, including general-purpose cores, vector processors, and GPGPUs. It will also extend the notion of quality programmability to the memory system and on-chip interconnect network. Techniques will be developed to optimize programs for execution on quality programmable platforms by identifying resilient computations, and tuning the degree to which they can be approximated while maintaining acceptable application-level output quality. The project will leverage outreach programs at Purdue, including Summer Undergraduate Research Fellowships (SURF), the NCN (Network for Computational Nanotechnology), and the Women and Minority in Engineering programs, to involve undergraduates and minority students in this research.

Agency
National Science Foundation (NSF)
Institute
Division of Computer and Network Systems (CNS)
Type
Standard Grant (Standard)
Application #
1423290
Program Officer
Marilyn McClure
Project Start
Project End
Budget Start
2014-10-01
Budget End
2018-09-30
Support Year
Fiscal Year
2014
Total Cost
$480,625
Indirect Cost
Name
Purdue University
Department
Type
DUNS #
City
West Lafayette
State
IN
Country
United States
Zip Code
47907