<title>Signal Processing Of Ultra Wideband Data Streams: Use Of Parallelism, New Device Technologies, And High System Clock Speeds</title>

Abstract

Signal processing of ultra wideband data streams:use of parallelism, new device technologies, and high system clock speedsB. K. Gilbert, T. M. Kinter, W. K. Van Nurden, and S. M. HartleySpecial- Purpose Processor Development Group, Mayo Foundation,200 First Street S.W., Rochester, Minnesota 55905R. ZuccaRockwell International Microelectronics Research and Development Center,1049 Camino Dos Rios, Thousand Oaks, California 91360AbstractArchitectural solutions to large signal and data processing problems can now in manycases be implemented in VLSI -complexity silicon integrated circuits in a majority ofcases, provided that the economic constraints of design costs, unit costs, and schedulecan be resolved. However, for a small subclass of signal processing tasks involving veryhigh input data bandwidths and /or large volume computation, VLSI implementations ofparallel architectures are a useful but incomplete solution, or in some cases even an ina-dequate one; alternate device technologies, e.g., high speed Gallium Arsenide digital com-ponents, must be considered for very high clock rate processors. This paper explores thetradeoffs between architectures and device technologies and the roles which they can playin very high bandwidth classical signal processing tasks.IntroductionA multitude of engineering research groups are presently exploring the methods bywhich the emerging very large scale integration (VLSI) technology, i.e., the ability toplace more than 10,000 logic gates on a single integrated circuit, can be exploited forthe solution of difficult data processing problems using parallel or distributed computerarchitectures. Many recent discussions of VLSI technology applications have ignored theproblems inherent in signal processing applications, which are typified by a set ofcharacteristics and constraints somewhat at variance with those of the data processingenvironment. We will concentrate in this presentation on a discussion of problems uniqueto signal processing, and try to assess the impact of the newest device technologies,including but not limited to silicon VLSI, on their solution.The characteristics of classical signal processing algorithms are quite differentfrom those of data processing algorithms, for the following reasons. Signal processingalgorithms generally operate on extremely long vectors of one, two, or more dimensions.The individual elements of these vectors are usually of relatively low precision,generally represented by fixed precision operands of eight to twelve bits in length, occa-sionally as complex fixed precision values in forms such as Real[12] + Imag[12], or evenless frequently, as block floating point operands or conventional low precision floatingpoint operands (1). Until recently, full precision floating point computation, with itsattendant hardware complexity, was rarely required, a situation which still prevails forthe majority of signal processing applications.Classical signal processing algorithms developed over the past three decades, andoften employed in a variety of combinations, include autocorrelation, cross correlation,convolution, fast Fourier transformation, multiplication of a vector by a scaler, for-mation of dot and cross products, matrix operations and binning (for synthetic apertureradar, doppler beam sharpening radar, and all forms of computed tomography) and Chirp -Ztransformation (for ultrasound computed tomography and in search radar applications). Allof these algorithms are characterized 1) by the regularity of the primitive arithmeticoperations, i.e., multiplications, additions, and subtractions, with which they areimplemented; 2) by a negligible amount of decision branching; and 3) by the large ratio ofcomputational steps to loop steps during their execution. Hence, a large number of globaloperations are implemented in a very similar manner. In addition, the use of two'scomplement fixed precision arithmetic reduces the addition and subtraction operations to asingle arithmetic type. Generalized fixed precision multiplication can often be imple-mented by combinations of binary shift operations, which are even easier to implement