Depends on the structure of your problem. Some lend themselves well to vectorization, others not nearly as much. Or at least the needed vectorized solvers are not yet available for some problem structures.

Try to use perf-tools to determine how you're currently using these execution units, and then you can look how much it might help. Rule of thumb: double the bitwidth of vector instructions gets you 80% more speed, instead of the theoretical 100%.

Those languages where the tensor/array is the fundamental unit: such as the APL family. J and K have optimized builds using AVX.

Not serve-y but RPCS3 benefits a lot from those.