Easy
Ten thousand 10-element launches versus one fused launch
Version A launches a tiny kernel that processes just 10 elements, 10,000 times in a host loop (100,000 elements total, all on the same stream). Version B launches one kernel with a grid-stride loop that processes all 100,000 elements in a single launch. Each individual element's compute work is trivial (a few FLOPs). Which is faster, and why?
19 of 46 in GPU / CUDA