Memory system design space exploration for low-power, real-time speech recognition

Authors:
Rajeev Krishna;Scott Mahlke;Todd Austin
Affiliations:
University of Michigan, Ann Arbor, MI;University of Michigan, Ann Arbor, MI;University of Michigan, Ann Arbor, MI
Venue:
Proceedings of the 2nd IEEE/ACM/IFIP international conference on Hardware/software codesign and system synthesis
Year:
2004

Citing 7
Cited 0

Power estimation for architectural exploration of HW/SW communication on system-level buses

CODES '99 Proceedings of the seventh international workshop on Hardware/software codesign
SimpleScalar: An Infrastructure for Computer System Modeling

Computer
Parallel Dispatch Queue: A Queue-Based Programming Abstraction To Parallelize Fine-Grain Communication Protocols

HPCA '99 Proceedings of the 5th International Symposium on High Performance Computer Architecture
A highly configurable cache architecture for embedded systems

Proceedings of the 30th annual international symposium on Computer architecture
A low-power accelerator for the SPHINX 3 speech recognition system

Proceedings of the 2003 international conference on Compilers, architecture and synthesis for embedded systems
Hardware architectures to support low power natural i/o applications

Hardware architectures to support low power natural i/o applications
A characterization of speech recognition on modern computer systems

WWC '01 Proceedings of the Workload Characterization, 2001. WWC-4. 2001 IEEE International Workshop

Quantified Score

Hi-index	0.00

Visualization

Abstract

The recent proliferation of computing technology has generated new interest natural I/O interface technologies such as speech recognition. Unfortunately, the computational and memory demands of such applications currently prohibit their use on low--power portable devices in anything more than their simplest forms. Previous work has demonstrated that the thread level concurrency inherent in this application domain can be used to dramatically improve performance with minimal impact on overall system energy consumption, but that such benefits are severely constrained by memory system bandwidth. This work presents a design space exploration of potential memory system architectures. A range of low--power memory organizations are considered, from conventional caching to more advanced system--on--chip implementations. We find that, given architectures able to exploit concurrency in this domain, large L2 based cache hierarchies and high bandwidth memory systems employing data stream partitioning and on--chip embedded DRAM and ROM technologies can provide much of the performance of idealized memory systems without violating the power constraints of the low--power domain.