Performance enhancement techniques for wavefront array filter structures

Systolic Arrays are a common implementation method for high performance pipelined DSP structures. In wavefront arrays, the global clock is removed and asynchronous handshaking communication is used to synchronise data transfers between Stages. One of the often claimed advantages of wavefront arrays is that data can be processed as soon as each stage is completed rather than being limited by the global synchronisation signal. Despite the theoretical foundations of wavefront arrays being well known for some time there has been little or no investigation Of their performance for particular algorithms in the presence of data dependant stage processing times. Using realistic assumptions for data dependant processing times in multipliers this paper reports on the performance of filters implemented as wavefront arrays with asynchronous communication protocols. The results show that wavefront arrays with data dependant processing times can achieve significantly higher throughput as compared with systolic arrays