Flash-KMeans: Fast and Memory-Efficient Exact K-Means
Summary
Flash-KMeans revisits K-means as an online primitive for modern GPU-based AI systems. It removes the IO bottleneck and atomic-write contention that limit existing GPU implementations by introducing FlashAssign and sort-inverse updates, along with co-designs for practicality, yielding substantial speedups over baselines like cuML and FAISS on NVIDIA H200 GPUs.