FPGA-Based Acceleration of Expectation Maximization Algorithm Using High-Level Synthesis

Abstract

Expectation Maximization (EM) is a soft clustering algorithm which partitions data iteratively into M clusters. It is one of the most popular data mining algorithms that uses Gaussian Mixture Models (GMM) for probability density modeling and is widely used in applications such as signal processing and Machine Learning (ML). EM requires high computation time when dealing with large data sets. This paper presents an optimized implementation of EM algorithm on Stratix V and Arria 10 FPGAs using Intel FPGA Software Development Kit (SDK) for Open Computing Language (OpenCL). Comparison of performance and power consumption between Central Processing Unit (CPU), Graphics Processing Unit (GPU) and FPGA is presented for various dimension and cluster sizes. Compared to Intel® Xeon® CPU E5-2637, our fully optimized OpenCL model for EM targeting Arria 10 FPGA achieved up to 1000x speedup in terms of throughput (T) and 5395x speedup in terms of throughput per unit of power consumed (T/P). Compared to previous research on EM-GMM implementation on GPUs, Arria 10 FPGA obtained up to 64.74x speedup (T) and 486.78x speedup (T/P).