First of all, please make postings on new subjects in a new topic.
It's very hard to estimate the precise number of atoms, since it depends very much on the geometry as well as numerical parameters. First of all, it differs between bulk and two-probe, and then the memory usage is quite sensitive to the number of k-points for instance, and of course basis set size. Generally, you can place a larger number of atoms in a "long" configuration (like a nanowire) than in a dense bulk systems.
Of course, in the end all depends on the amount of memory you have available too! With 8 Gb RAM you can do quite many more atoms than with 2 Gb.
I prefer not to comment on ATK 2008.02 since it has been replaced by ATK 2008.10, and one of the main points of that release was to improve memory handling to allow larger systems.
For example, we have done a MgO supercell with ATK 2008.10 with 1024 atoms, using a good quality basis set (double-zeta polarized, DZP). It required about 3 Gb of RAM and took 16 hours to compute on a single node.
A Si/Au two-probe with 1,166 atoms in the equivalent bulk took 24 hours on 10 nodes, using 2 Gb with a single-zeta basis set.
In both cases there was no k-point sampling in the XY directions; the unit cell is so large that it shouldn't really be necessary. Both cases are also un-polarized. For magnetic atoms, with spin, a rough estimate is half the capacity.
