Energy efficient calculations of text similarity measure on FPGA-accelerated computing platforms

Abstract

This paper presents an impact of the customized hardware accelerator on the overall performance of the text similarity computing system. The hardware processing module that is presented in the paper is a building block of the processing engine in the search system of related documents. The engine is used in the phase of preliminary retrieval of similar documents. The TF-IDF weighting scheme and cosine similarity metric are used by the module. Evaluation boards equipped with Xilinx’s Field Programmable Gate Array (FPGA) were utilized as a hardware platforms for implementation of the selected time-consuming operations. The series of tests was conducted, and the results of the hardware-accelerated solutions were compared against the standard software implementation. The two different FPGA-enabled platforms were employed in the experiments. The low-power and the high-performance platform were used to compare the metrics of different hardware solutions. We provide the adequate results and conclusions that present that the energy and speed metrics of the text similarity calculations can be improved thanks to the hardware accelerator. Consequently, the cluster of FPGA-enabled nodes is proposed for the large scale processing.