A two-probe system is periodic in the two directions perpendicular to the transport direction (which we always take as Z for simplicity). Therefore, it's easy to set up a bulk electrode cell to represent a whole metal surface, for instance. You just have to remember to make the surface unit cell large enough that the molecule in the central region does not interact with it's repeated copies.
If, on the other hand, you wish to have a 1D-type of electrode, like a nanotube, you have the opposite "problem"; you must include enough vacuum in the XY unit cell that the electrodes have no interactions with their repeated copies.
Compare the two attached figures, one for a bulk electrode ([111] 3x3 fcc Au) and one for a nanotube. Both two-probe systems are repeated 3x2 times in the transverse directions; note the unit cells!
I'm not sure I understood the question on hardware, but regarding the vacuum, obviously this adds to both the CPU and memory usage, since the real-space 3D grids can be quite large, and the Poisson equation will be heavier to solve. Usually this is, however, not a problematic constraint that limits the size of the calculation.
1D systems typically exhibit a slightly worse performance gain in parallel due to the 1x1 k-point sampling in the XY plane. On the other hand, the way ATK handles memory partitions along the transport axis, and the reduced number of neighbor atoms, typically means that you can handle a larger number of atoms in such an "elongated" configuration compared to a compact system.
