Accelerating k-Means on GPU with CUDA Programming. Issue 1 (March 2020)

• Record Type: Journal Article
• Authors: Yang, Can; Li, Yin; Cheng, Fenhua
• Is Part Of: IOP conference series. Volume 790:Issue 1(2020)
• Publication Date: 2020-03
• Subjects: k-Means -- GPU -- CUDA; Materials science -- Periodicals; 620.1105
• DOI: 10.1088/1757-899X/790/1/012036
• Languages: English
• ISSNs: 1757-8981

Abstract

Abstract: We accelerate basic k-Means algorithm using CUDA GPU, a new programming model by NVIDIA, and experiment data shows we achieve a maximum speedup of 67.752, while other teams claim 20 to 40. Also we find that the basic k-Means algorithm is most sensitive to the cluster size k, and less to the datasets size b and least to the dimension d. In addition, we find the CUDA shared memory improves the performance, but also depends on which factor we scale.