Scalable and parameterised VLSI architecture for efficient sparse approximation in FPGAs and SoCs

摘要

A parameterised and scalable very large scale integration (VLSI) soft intellectual property (IP) is presented that can be implemented in programmable logic devices, such as field programmable gate arrays (FPGAs) or a system-on-chip design for efficient sparse approximation. The proposed architecture is optimised based on the orthogonal matching pursuit algorithm by both algorithm reformulation and architecture resource sharing techniques. The soft IP core supports a floating-point data format with 10 design parameters, which provides the necessary flexibility for application-specific customisation. The soft IP is evaluated on various FPGA platforms. The evaluation results show that design can achieve up to 30% higher throughput than the existing solutions while offering a larger dynamic range capability and better design flexibility.