M07 Low Power: Fixed-point refinement, a guaranteed approach towards energy efficient computing

Printer-friendly versionPDF version

Agenda

TimeLabelSession
09:30M07.1Fixed-point refinement, a guaranteed approach towards energy efficient computing: Part I

Organiser:
Olivier Sentieys, INRIA - University of Rennes 1, FR

09:30M07.1.11. Introduction
Olivier Sentieys, INRIA - University of Rennes 1, FR

  • What is fixed-point arithmetic?
  • Why to use fixed-point?
09:45M07.1.22. Fixed-point arithmetic
Olivier Sentieys, INRIA - University of Rennes 1, FR

  • Arithmetic rules and type support in languages and architectures
  • Review of commercial and academic tools
  • Precision metrics: average noise vs. bounds
10:15M07.1.33. Finite word-length effect analysis
Daniel Menard, INSA Rennes/IETR, FR

  • Guaranteed range evaluation
  • Dealing with overflows: stochastic approaches
  • Numerical accuracy evaluation: bounded or average errors
  • Dealing with system-level complexity
11:30M07.2Fixed-point refinement, a guaranteed approach towards energy efficient computing: Part II

Organiser:
Olivier Sentieys, INRIA - University of Rennes 1, FR

11:30M07.2.14. High-level synthesis under accuracy constraint
Karthick Parashar, IMEC, BE

  • Word-length optimization: an iterative and NP-hard problem
  • The intricacy of joint HLS and WLO
  • System-level optimization: dealing with complexity
12:15M07.2.25. SIMD vectorization under accuracy constraint
David Novo, EPFL, CH

  • Soft SIMD vs. Hard SIMD
  • Exploring vectorization vs. accuracy trade-off
  • Results: HEVC video processing and video surveillance
12:35M07.2.36. Opportunistic run-time precision adaptation
David Novo, EPFL, CH

  • Dynamic adaption of processing accuracy
  • When flexibility leads to higher energy efficiency
12:50M07.2.47. Conclusion and open research directions
David Novo, EPFL, CH