The latter. For most uses of Pandas you can simply use native arrays. You can use for loops and so on.

For N-dimensional array computation: https://github.com/mratsim/Arraymancer/