I wrote two approaches to the same problem.
The first approach uses simple python data structures and greedy evaluation. It runs under CPython in 0.15 seconds. Running under pypy takes 1.2 seconds. pypy is 8x slower.
The second approach (using the same data) builds a big graph and visits nodes v^3 times. Running under CPython takes 4.5 seconds. Running under pypy takes 1.6 seconds. pypy is almost 3x faster.
So... that's why. "It depends." But—it's great we have two implementations of one language where one jits repetitive operations and the other evaluates straight-through code faster.