Design Trade-offs in Configurable FPGA Architectures for K-Means Clustering
Alexandru Amaricăi · Studies in Informatics and Control · 2017
K-Means clustering is a popular technique for data partitioning, frequently used in data mining.The simple control flow, and high degree of parallelism, makes it a good candidate for FPGA acceleration.We propose a highly configurable architecture, based on Euclidean distance computation.It can be tuned by the following parameters: number of dimensions, dimension width, dimension based parallelism degree, number of centroids and centroid based parallelism degree.We study their impact on different K-Means components, such as the distance computation, distance comparison, accumulation, division, or the memory modules within the accelerator.Furthermore, for the aforementioned parameters we investigate the performance/cost trade-offs of the proposed K-Means accelerator implementation.