Power-mode-aware Memory Subsystem Optimization for Low-power System-on-Chip Design

The memory subsystem is increasingly subject to an intensive energy minimization effort in embedded and System-on-Chip development. While the main focus is typically put on energy consumption reduction, there are other optimization aspects that become more and more relevant as well, e.g., peak power...

Full description

Saved in:
Bibliographic Details
Published inACM transactions on embedded computing systems Vol. 18; no. 5; pp. 1 - 25
Main Authors Strobel, Manuel, Radetzki, Martin
Format Journal Article
LanguageEnglish
Published 01.10.2019
Online AccessGet full text

Cover

Loading…
More Information
Summary:The memory subsystem is increasingly subject to an intensive energy minimization effort in embedded and System-on-Chip development. While the main focus is typically put on energy consumption reduction, there are other optimization aspects that become more and more relevant as well, e.g., peak power constraints or time budgets. In this regard, the present article makes the following contributions. Taking industrial-grade information into account, different Static Random-Access Memory (SRAM) power modes and their characteristics are presented at first. Using this information, a comprehensive optimization model with the main intention of energy minimization is defined. It is based on memory access statistics that represent the embedded software of interest, which allows for application-tailored improvements. Further, it considers different power states of the memory subsystem and enables the definition of peak power and time corridor constraints. The presented two-stage implementation of this optimization model allows the handling of large design spaces. Clearly defined interfaces facilitate the exchange of individual workflow parts in a plug-and-play fashion and further enable a neat integration of our optimization method with existing hardware/software (HW/SW) codesign synthesis flows. A general evaluation for different technology nodes yields that the optimization potential of memory low-power modes increases with advancing miniaturization but also depends on the data footprint of the embedded software. Experimental results for a set of benchmark applications confirm these findings and provide energy savings of up to 90% and over 60% on average compared to a monolithic memory layout without low-power modes.
ISSN:1539-9087
1558-3465
DOI:10.1145/3356583