https://github.com/ragibson/cuda-k-means

An implementation of Lloyd's algorithm for data clustering on GPUs and computational accelerators.
https://github.com/ragibson/cuda-k-means

clustering cuda gpu k-means unsupervised-clustering

Last synced: 2 months ago
JSON representation

An implementation of Lloyd's algorithm for data clustering on GPUs and computational accelerators.

• Host: GitHub
• URL: https://github.com/ragibson/cuda-k-means
• Owner: ragibson
• License: mit
• Created: 2019-05-22T18:12:37.000Z (almost 6 years ago)
• Default Branch: master
• Last Pushed: 2019-05-23T19:44:12.000Z (almost 6 years ago)
• Last Synced: 2025-01-05T19:42:36.562Z (4 months ago)
• Topics: clustering, cuda, gpu, k-means, unsupervised-clustering
• Language: TeX
• Homepage:
• Size: 1.15 MB
• Stars: 3
• Watchers: 3
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# CUDA-k-means

This repository contains three main implementation of Lloyd's k-means

algorithm:

* [kmeans_cpu.c](kmeans_cpu.c): A k-means implementation for use on a single

CPU core.

* [kmeans_gpu.cu](kmeans_gpu.cu): A k-means implementation for use on CUDA

GPUs.

* [kmeans_atomic.cu](kmeans_atomic.cu): A k-means implementation for use on

CUDA GPUs, using atomic operations (for comparison purposes).

In terms of "number of cluster updates per second", these programs

peak at

* [kmeans_cpu.c](kmeans_cpu.c): ~950 million computations per second

* [kmeans_gpu.cu](kmeans_gpu.cu): ~65 billion computations per second

* [kmeans_atomic.cu](kmeans_atomic.cu): ~575 billion computations per second

(though this is only feasible for a large number of clusters)

on a system with an E5-2650v3 CPU and TITAN V GPU.

A (significantly) more detailed performance analysis can be read

[here](report/report.pdf).