SIMD Architectures for Radar Signal Processing and Artificial Neural Networks
The thesis is about computer architectures specially tuned to an application area. This means that the work spans the area from implementation technology via processor and computer system organization to the applications themselves. The work reported here is in the area of embedded high performance computing, near the area of application specific hardware. The thread throughout the thesis is how to design computers to suit a specific application area, while maintaining as much computing performance, programmability, scalability and flexibility as possible. The idea is that the multiple SIMD computing model can be a flexible and reasonably scalable concept for the high end applications considered. To test this hypothesis the approach taken is to use application examples, algorithm analysis and implementation experiments to derive suitable computing modules. These modules are then evaluated according to scalability, generality, efficiency, and implementation aspects. The application areas are artificial neural network computing and signal processing in phased array radar. For the artificial neural network computing a multiple SIMD architecture is suggested and artificial neural network algorithms are mapped onto a typical such module. Implementation aspects are discussed and the design of a prototype is shown. Then the use of artificial neural networks in an industrial real-time application is presented. The artificial neural networks are used to extract information from noisy and non-linear signals in combustion engines. It is shown that the neural networks are feasible, and close to optimal, in this application. In the area of signal processing for phased array radar, two application examples are analyzed and architectures suitable for the these are derived. An intermodule communication for implementation on a fiber-optic network is evaluated in a radar application. Then implementation issues for the processing modules are considered and discussed. This is done in the light of instruction statistics gathered from the application examples. Finally, the results are combined and the VEGA architecture is described and motivated.
In the thesis it is shown that the modular, multiple SIMD model can be efficiently used in both signal processing for phased array radar and artificial neural network computing. Furthermore, a conclusion drawn is that the linear array SIMD module with broadcast and ring communication is enough for many popular neural network models. It is also concluded that the moderately parallel MIMD machine with moderately parallel SIMD computing modules is a feasible architecture for signal processing in phased array radar.