This kind of side-by-side benchmarking is exactly what compiler and library engineers do before deciding a SIMD rewrite is worth the added code complexity, and cloud providers like AWS explicitly document which instance types guarantee AVX2 so software can target it as a safe baseline. The lesson that speedup depends heavily on whether a workload is memory-bound or compute-bound is the same reasoning Intel's own optimization manual walks through before recommending any vectorization effort.