A group project exploring GPU acceleration for k-nearest neighbours on the MNIST dataset.
The implementation includes kernels for Euclidean distances and bitonic-sort neighbour
selection, and uses CUDA streams, pinned host memory, and float4 loads, with per-stage
timing through CUDA events.
The repository reports 96.65% accuracy on the 10,000-image test set. The site labels this as a group project and does not attribute the team’s result to one person.