Pages 4 and 5 of this article goes more in depth about the differences between using KD-trees and BVH's (how the cost functions would differ with regards to the author's implementation as well as different ways to optimally partition): https://graphics.stanford.edu/~boulos/papers/togbvh.pdf