SIMD Reveals Limits For K-Means Vectorization

An author investigates SIMD (AVX-512) performance for K-Means image segmentation, benchmarking scalar, auto-vectorized, and hand-written intrinsics on an AMD EPYC 9654. Using a 5 million–pixel dataset, K=8, 20 iterations and ~20 GFLOPs total, the best compilers delivered 1.4s versus a theoretical 337ms peak, revealing large gaps and favoring intrinsics or CUDA for practical speedups.