Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{205736,
        author = {TEJASVI P C and Sowmya Sunkara},
        title = {Resource-Efficient Streaming FPGA Accelerator for Generalized K-Means Clustering with Comparator-Driven Early Pruning and Flush-Controlled Pipeline Architecture},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {1},
        pages = {8254-8263},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=205736},
        abstract = {K-means clustering is a fundamental unsupervised machine learning primitive with applications spanning medical image segmentation, network intrusion detection, convolutional feature learning, and AV1 video codec palette mode processing. Existing field-programmable gate array (FPGA) accelerators impose prohibitive digital signal processor (DSP) costs through spatial parallelism, or sacrifice dataset generality through application-specific co-optimization. This paper presents a resource-efficient streaming FPGA accelerator for generalized K-means clustering on the Microchip PolarFire MPF300T FPGA, incorporating three co-designed innovations: (1) a dimension-serialized streaming distance-computation pipeline achieving O(d) DSP utilization through temporal hardware reuse; (2) a memory-free comparator-chain early-pruning mechanism with O(k) register cost versus O(n·k) for triangle-inequality filtering; and (3) a three-state flush-controlled finite state machine (FSM) guaranteeing zero stale-data propagation across centroid traversals. Post-layout evaluation across six UCI Machine Learning Repository benchmark datasets (d = 3–28, k = 38–82) demonstrates 151–153 MHz operation at 45,891–65,796 look-up tables (LUTs), 6–8 DSP blocks, and 128–138 mW of total power, achieving a 216× DSP reduction over a state-of-the-art AV1-oriented accelerator at competitive frequency and full dataset generality.},
        keywords = {Comparator-driven pruning, edge AI accelerator, flush-controlled finite state machine, FPGA implementation, generalized K-means clustering, PolarFire FPGA, resource-efficient hardware, streaming pipeline.},
        month = {June},
        }