(author here) Both would work - doesn't fundamentally change the underlying idea. This particular representation is lifted from raytracing in one weekend and emphasizes the fact that we're really thinking of aabb as an intersection of three "slabs" of space. for me, it makes the ray intersection bit somewhat easier to explain.

SIMD and data locality. You probably want to check across three vectors simultaneously and load the coordinates next to each other.

I'm guessing here. I haven't written video games in 20 years but struct packing/alignment was super important on the Sony PSP back then.

For SIMD at least, the {mins[3], maxs[3]} representation aligns more naturally with actual instructions on x86. To compute a new bounding box:

new_box.mins = _mm_min_ps(a.mins[3], b.mins[3]);

You would want [4] not [3], with the last one being padding. Of course, you can't always afford that.

Indeed. This is classic array-of-structs versus struct-of-arrays.