Inconsistent execution time of tiled matrix multiplication algorithm (CUDA)

Question

I\'ve been trying to see for myself the execution time difference between a normal non-tiled matrix multiplication algorithm in CUDA and a tiled one. However, I don\'t under