A systolic VSLI chip for implementing orthogonal transforms

Abstract

The design of a systolic VLSI chip is described for the implementing of signal processing algorithms that may be decomposed into a product of simple real rotations. These include orthogonal transformations. Applications of this chip include projections, discrete Fourier and cosine transforms, and geometrical transformations. Large transforms may be computed by 'tilting' together many chips for increased throughput. A CMOS VLSI chip containing 138000 transistors in a 5*3 array of rotators has been designed, fabricated, and tested. The chip has a 32-MHz clock and performs real rotations at a rate of 30 MHz. The systolic nature of the chip makes use of a fully synchronous bit-serial interconnect and a very regular structure at the rotator and bit levels. A distributed arithmetic scheme is used to implement the matrix-vector multiplication of the rotation. >