A VLSI systolic array processor for complex singular value decomposition

Abstract

The singular value decomposition is one example of a variety of more complex routines that are finding use in modern high performance signal processing systems. In the interest of achieving the maximum possible performance, a systolic array processor for computing the singular value decomposition of an arbitrary complex matrix was designed using a silicon compiler system. This system allows for ease of design by specification of the processor architecture in a high level language, utilizing parts from a variety of cell libraries, while still benefiting from the power of custom VLSI. The level of abstraction provided by this system allowed more complex functional units to be built up from existing simple library parts. A novel fast interpolation cell for computation of square roots and inverse square roots was designed, allowing for a new algebraic approach to the singular value decomposition problem. The processors connect together in a systolic array to maximize computational efficiency while minimizing overhead due to high communication requirements.