This paper explores the effectiveness of different knobs to dynamically trade energy consumption with output quality in approximate SRAMs for error-tolerant applications (such as video). Leveraging the different impact of errors on quality at most significant bit (MSB) and least significant bit (LSB) positions, energy savings higher than those provided by simple voltage scaling are enabled. Firstly, a comparison of two techniques, dual-VDD and LSB dropping, is carried out showing that the latter is preferable thanks to its intrinsic simplicity and more pronounced energy savings. Secondly, a selective Error Correction Code (ECC) technique which reuses the LSBs as check bits to protect MSBs is investigated. Measurements on a 28nm CMOS 32kb SRAM show that bit dropping and bit reuse achieve an energy reduction of up to 33% and 28%, compared to simple voltage scaling at iso-quality. When combined together, the two techniques achieve a better energy saving (40%) and a supply voltage reduction of about 100mV at iso-quality. Finally, guidelines to select the energy-optimal combination of the two techniques are provided for a given quality target.

Better-than-voltage scaling energy reduction in approximate SRAMs via bit dropping and bit reuse

Frustaci F.
;
Alioto M.
2015-01-01

Abstract

This paper explores the effectiveness of different knobs to dynamically trade energy consumption with output quality in approximate SRAMs for error-tolerant applications (such as video). Leveraging the different impact of errors on quality at most significant bit (MSB) and least significant bit (LSB) positions, energy savings higher than those provided by simple voltage scaling are enabled. Firstly, a comparison of two techniques, dual-VDD and LSB dropping, is carried out showing that the latter is preferable thanks to its intrinsic simplicity and more pronounced energy savings. Secondly, a selective Error Correction Code (ECC) technique which reuses the LSBs as check bits to protect MSBs is investigated. Measurements on a 28nm CMOS 32kb SRAM show that bit dropping and bit reuse achieve an energy reduction of up to 33% and 28%, compared to simple voltage scaling at iso-quality. When combined together, the two techniques achieve a better energy saving (40%) and a supply voltage reduction of about 100mV at iso-quality. Finally, guidelines to select the energy-optimal combination of the two techniques are provided for a given quality target.
2015
9781467394192
approximate computing; energy-quality tradeoff; Error-tolerant; SRAM; ultra-low power processing; Hardware and Architecture; Electrical and Electronic Engineering
File in questo prodotto:
Non ci sono file associati a questo prodotto.

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.11770/287197
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo

Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 13
  • ???jsp.display-item.citation.isi??? 11
social impact