PROGRAMMABLE TECHNOLOGIES WEB SITE

A scientific study of the problems of digital engineering for space flight systems,
with a view to their practical solution.


Paper A2 Synopsis

Sample Integrated Fourier Transform (SIFT) With High-Performance ASIC Implementation

2001 MAPLD International Conference

Kossiakoff Conference Center
The Johns Hopkins University - Applied Physics Laboratory
11100 Johns Hopkins Road
Laurel, Maryland 20723-6099

September 11-13, 2001


Walter E. Pelton
Faster Fourier Transforms, Inc.
3584 Lancelot Ct. 
Fremont, CA 94536
walter@alumni.caltech.edu  

Trang K. Ta
Fujitsu Microelectronics, Inc.
3545 1 st St. N
Santa Clara, CA 95134
tta@fmi.fujitsu.com

Nipa Yossakda 
NPU
117 Fourier Ave.
Fremont, CA
nipayoss@yahoo.com

Pochang Hsu
NPU
117 Fourier Ave.
Fremont, CA
phsu@pulsent.com

Abstract

A zero-latency Fourier Transform algorithm and a high-performance ASIC implementation are presented. A 64-point complex Fourier Transform ASIC was simulated; results: 8mW, .49mm2, 12.8µsec. A 1024-point Fourier Transform is proposed: 500mW, 32mm2, 3.2µsec.

Table of Contents

  1. Introduction
  2. Overview
    1. DFT and FFT
    2. Butterfly and Architecture
  3. SIFT Paradigm
  4. Implementation And Results
  5. Conclusions

List of Figures

Figure 1: FFT vs. SIFT data pipelines
Figure 2: Number of cycles vs. points for SIFT, single precision FFT(s), and double precision FFT(d)
Figure 3: SIFT: execution as samples arrive
Figure 4: Block diagram of the 64 point SIFT
Figure 5: Each coefficient accumulated in parallel
Figure 6: Origin offset does not alter waveform
Figure 7: Micro-graph of 64-point FT core
Figure 8: Architecture of 1024-complex-point FT
Figure 9 : Main portion of the Aspect Generator
Figure 10: FT of sin(x)
Figure 11: FFT Butterfly storage, none for SIFT
Figure 12: Inverse FT of Acos(2pf1 x/N) + Bsin(2pf2 x/N) at frequencies (16,63)
Figure13: Image of 64-point SIFT PCB
Figure 14: Inverse FT of Ecos(2pf1 x/N) + Fsin(2pf2 x/N) at frequencies (31,63)

List of Tables

Table 1. # MAC cycles vs. # Points in Transform

Summary

A new Fourier Transform process and architecture were presented which simplify hardware and support parallelism without latency. Reduced power, clock cycles and cycle times were shown. A method to further improve the computational efficiency was proposed.


Home
January 09, 2002
Digital Engineering Institute
Web Designer: Richard Katz