[flostk]

In this interview Mike Clark from AMD explains that a little, search for "third scheduler":

https://chipsandcheese.com/2024/07/15/a-video-interview-with...

The way I understand it, it's a combination of a unified scheduler for floating point being difficult to implement because many FP instructions need multiple cycles, and FP code being more regular in practice, so you don't need the scheduler to be as powerful.