ECube: An Efficient Architecture for Analyzing Time-Varying Spectra
Mohan Vishwanath, Robert Michael Owens, M.J. Irwin · 2005
In this paper we describe ECube, a highly regular semi-systolic architecture for computing timefrequency distributions. This is an enhanced version of the Arithmetic Cube II which is currently being built at Penn Slate. The Discrete Short Time Fourier Transform and the Discrele Wavelet Transform, the most general tools for the analysis of time-varying spectra, can be computed in an optimal manner on the ECube. This architecture is based on the small n algorithms and their formulations by Winograd and Good. ECube does very few multiplications at the expense of Q large number of additions. It uses an AT2 optimal adder array, based on maximally redundant radix 4, to do the additions. A dual pipeline scheme is used to reduce the overheads due lo pre and post multiplications.