Processor architectures for synthetic aperture radar

This thesis examines processor architectures for Synthetic Aperture Radar (SAR). SAR is a remote sensing technique that requires large amounts of computation and memory to form images. Processor architectures are sought that exhibit high performance, are scalable, are flexible, and are cost effective to develop and build. Performance is taken to be the primary figure of merit. The three facets of systems design, namely algorithm, technology, and architecture, are each examined in the process of finding the best architecture implementations. The examination of the algorithms is begun by reviewing SAR processing theory with the intent of summarizing the background for typical SAR processor performance requirements. A representative set of SAR algorithms is analyzed to determine and compare their computational requirements and to characterize them in terms of basic digital signal processing (DSP) operation types. The algorithm partitioning options for parallel processing are classified and compared. The area of technology is addressed by surveying computing technologies that are relevant to SAR processing. The technologies considered cover computational devices, memory devices and interconnections. The knowledge of SAR algorithms and computing technology are then combined to study design considerations for the memory and interconnection subsystems of SAR processors. The requirement for two dimensional access to large data arrays is found to be the main complicating factor in memory design. The judicious use of wide data path widths, caches, interleaving and fast memory is discussed as a solution to the memory latency problem. The applicability of the most commonly used interconnection networks is examined. Buses, meshes, and crossbars are all found to be effective in certain situations. A classification of architectural approaches adapted to describe current and future SAR processors is used as the framework for the architecture selection. Feasible implementations of the architectural approaches are identified and their suitability for SAR is analyzed. Three approaches are identified as the most promising: networks of workstations, DSP chips, and field programmable gate arrays (FPGAs). A detailed examination is made of these three approaches. Four variations of the DSP approach are considered: general purpose DSPs, vector DSPs, a proposed optimised DSP, and digital filter devices. Each approach offers a different trade-off between performance and flexibility. The most cost effective architectures approaches are found to be those based on general purpose or vector DSPs. An original heterogeneous design is presented that combines the strengths of these two approaches.Applied Science, Faculty ofElectrical and Computer Engineering, Department ofGraduat