Stochastic Number Duplicators Based on Bit Re-Arrangement Using Randomized Bit Streams

Ryota ISHIKAWA  Masashi TAWADA  Masao YANAGISAWA  Nozomu TOGAWA  

Publication
IEICE TRANSACTIONS on Fundamentals of Electronics, Communications and Computer Sciences   Vol.E101-A   No.7   pp.1002-1013
Publication Date: 2018/07/01
Online ISSN: 1745-1337
DOI: 10.1587/transfun.E101.A.1002
Type of Manuscript: Special Section PAPER (Special Section on Design Methodologies for System on a Chip)
Category: 
Keyword: 
stochastic computing,  stochastic number,  duplicator,  bit re-arrangement,  re-convergence path,  

Full Text: PDF(1.5MB)
>>Buy this Article


Summary: 
Recently, stochastic computing based on stochastic numbers attracts attention as an effective computation method, which realizes arithmetic operations by simple logic circuits with a tolerance of bit errors. When we input two or more identical values to a stochastic circuit, we require to duplicate a stochastic number. However, if bit streams of duplicated stochastic numbers are dependent on each other, their arithmetic operation results can be inaccurate. In this paper, we propose two stochastic number duplicators, called FSR and RRR. The stochastic numbers duplicated by the FSR and RRR duplicators have the equivalent values but have independent bit streams, effectively utilizing bit re-arrangement using randomized bit streams. Experimental evaluation results demonstrate that the RRR duplicator, in particular, obtains more accurate results even if a circuit has re-convergence paths, reducing the mean square errors by 20%-89% compared to a conventional stochastic number duplicator.