Channel Filter Cross-Layer Optimization
MetadataVis full innførsel
The recent raise of Internet of Things has increased the demand of energy-efficient wireless devices. However, the design process of a low-energy, high-performance device for all operational cases is not trivial. Thus, in this thesis a cross layer optimization technique called algorithm-architecture co-design is used to optimize one of the most critical DSP blocks in any communication device, namely, the channel filter. In order to effectively trade between similar RTL designs in terms of area and power dissipation, a fully automated tool-flow is created which performs RTL-simulation, synthesis, layout and power analysis. The tool-flow provides the results of a 500 gate design in less than 5 minutes, running on a computer of the current industry standard, and is considered to be very accurate based on the results of a previous study. A digital low pass filter is first optimized through a constructed filter sorting algorithm. It generates a large number of theoretical filter solutions and sorts them based on how eligible they are for hardware implementation. The algorithm is made generic, and hence applicable for any filter requirement, and proves to find the most energy-efficient solution. The hardware architecture of the most effective filter implementation is then thoroughly analyzed in two stages. Firstly, in order to find the filter's most effective quantization levels, and secondly, in order to make the architecture dynamic with regards to filter performance and power dissipation. In the latter, two main approaches are proposed. The first approach adapts the filter order, and hence the stopband attenuation, according to the quality of the radio link. The best implementation of this approach manages to reduce the power dissipation of 28%, 55% and 88% for the constructed low power modes, with an increase of only 8% in area compared to the non-dynamic implementation. The second approach adapts the quantization level, and hence the amount of noise introduced by the filter, according to the radio link. Here, the best implementation reduces the power dissipation of 11%, 32% and 81% for the low power modes, while increasing the area of only 18%.