All your cores, the same answer.

Any two systems that touch the same data are ordered, or proven not to collide; every floating-point sum is added in a fixed order. No speed-up figure is claimed. The full page is being written; the tour shows it today.