[Colloquium] Reminder: Lee Ehudin/MS Presentation - May 25, 2017

[Karin Czaplewski via Colloquium]

This is a reminder of Lee Ehudin’s MS Presentation.

===========================
Department of Computer Science
The University of Chicago

Date: Thursday, May 25, 2017

Time: 10:30am

Location: Young 302 <https://maps.uchicago.edu/?location=Young+Memorial+Building> (5555 S. Ellis Avenue)

Bx/MS Candidate: Lee Ehudin

MS Advisor:  Hank Hoffman

MS Paper Title:  NEAT: A Tool for Automated Exploration of Approximate FPU Designs

Abstract:

Much recent research is devoted to exploring tradeoffs between computational accuracy and energy. In particular, a number of techniques have been proposed for producing and using approximate arithmetic units that return an inexact answer with greatly reduced energy consumption. As the number of approximate techniques increases, the options for creating approximate programs explodes, creating the need for tools that help programmers explore the effects of approximation and combine different approximation techniques to achieve the lowest energy consumption for an accuracy constraint or the best accuracy for an energy constraint. To address this need, we present NEAT: a PIN tool that automatically explores the accuracy-energy tradeoff space for floating-point computation. NEAT accepts one or more user-defined approximate floating-point implementations and rules for when to substitute different implementations. NEAT then computes the floating-point operations in an application using those implementations and rules. We evaluate NEAT through a case study on 8 different applications and compare a set of rules that allows only one floating-point implementation per program to a set of rules that allow one approximation per function. We find that more of the accuracy-energy design space can be explored with the per-function rules than the single floating-point implementation. We also find that data collected from smaller inputs using both sets of rules is highly correlated to data collected from moderately-sized inputs.

A copy of Lee’s MS paper is attached.






Karin Czaplewski
Student & Faculty Services Specialist
Masters Program in Computer Science
The University of Chicago
5555 S. Ellis Avenue - Room 305
(773) 834-8587
Email: karin at cs.uchicago.edu <mailto:karin at cs.uchicago.edu>
http://csmasters.uchicago.edu <http://csmasters.uchicago.edu/>
Facebook.com/UChicagoMastersCS <http://facebook.com/UChicagoMastersCS>
Twitter.com/UChicagoMPCS <http://twitter.com/UChicagoMPCS>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20170523/449028e5/attachment-0002.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: thesis.pdf
Type: application/pdf
Size: 501779 bytes
Desc: not available
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20170523/449028e5/attachment-0001.pdf>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20170523/449028e5/attachment-0003.html>