CUDA k-means · interactive execution model

Runs the actual algorithm from danielberhane/cuda-kmeans on the real benchmark dataset, in your browser, and draws what the GPU does with it. Change the parameters and everything below recomputes from scratch.

Clusters (K) 4

Threads / block 128

Threshold 0.0010

Dataset

benchmark clustered

Points

Coloured by cluster. The benchmark data has no clusters in it, so k-means partitions it into Voronoi cells rather than recovering structure — switch to clustered to see the difference. The vertical banding is real: dimension 1 draws from only 19,968 distinct values, and 288 of them account for 80% of the points.

Plot dimensions

Grid · one thread per point

Each cell is one thread block, shaded by how many of its points changed cluster this iteration.

Shared memory per block

One extern __shared__ allocation, sliced into six regions. Drawn to scale.

State

Iteration

–

δ changed/N

–

Points moved

–

Grid

–

Reduce kernel

–

Shared / block

–

Dataset

–

Trace computed in

–

Cluster sizes

–