Skip to content

Latest commit

 

History

History
33 lines (20 loc) · 1.07 KB

An Ultra Low-Power Hardware Accelerator for Automatic Speech Recognition.md

File metadata and controls

33 lines (20 loc) · 1.07 KB
Raw

Paper title:

An Ultra Low-Power Hardware Accelerator for Automatic Speech Recognition

Publication:

MICRO’16

Problem to solve:

  1. Viterbi search in ASR (Automatic Speech Recognition) takes dominant part on time consumption.

  2. Using GPU to speed up this process needs high energy consumption.

As a result, in mobile devices, an accelerator for Viterbi search is required.

Major contribution:

  1. An accelerator for the Viterbi search that achieves 1.7x speedup over a high-end desktop GPU (980), while consuming 287x less energy.

  2. A prefetching architecture tailored to the characteristics of the Viterbi search algorithm that provides 1.87x speedup over the base design.

  3. A memory bandwidth saving technique that removes 20% of the accesses to off-chip system memory.

Lessons learnt:

This paper focus on only one algorithm in speech recognition, which is important and useful. The author finds out the bottleneck of the algorithm is the memory load part, and this leads to the whole accelerator design.