Energy area and speed optimized signal processing on FPGA

Abstract

Matrix multiplication and Fast Fourier transform are two computational intensive DSP functions widely used as kernel operations in the applications such as graphics, imaging and wireless communication. Traditionally the performance metrics for signal processing has been latency and throughput. Energy efficiency has become increasingly important with proliferation of portable mobile devices as in software defined radio. A FPGA based system is a viable solution for requirement of adaptability and high computational power. But one limitation in FPGA is the limitation of resources. So there is need for optimization between energy, area and latency. There are numerous ways to map an algorithm to FPGA. So for the process of optimization the parameters must be determined by low level simulation of each of the designs possible which gives rise to vast time consumption. So there is need for a high level energy model in which parameters can be determined at algorithm and architectural level rather than low level simulation. In this dissertation matrix multiplication algorithms are implemented with pipelining and parallel processing features to increase throughput and reduce latency there by reduce the energy dissipation. But it increases area by the increased numbers of processing elements. The major area of the design is used by multiplier which further increases with increase in input word width which is difficult for VLSI implementation. So a word width decomposition technique is used with these algorithms to keep the size of multipliers fixed irrespective of the width of input data. FFT algorithms are implemented with pipelining to increase throughput. To reduce energy and area due to the complex multipliers used in the design for multiplication with twiddle factors, distributed arithmetic is used to provide multiplier less architecture. To compensate speed performance parallel distributed arithmetic models are used. This dissertation also proposes method of optimization of the parameters at high level for these two kernel applications by constructing a high level energy model using specified algorithms and architectures. Results obtained from the model are compared with those obtained from low level simulation for estimation of error.