@DickH: this of course depends on the application. For memory and logic, 3 die stacks are quite common and I have also come across 5 dice stacked. You may want to look up Micron's Hypercube.
The author correctly points out the likely scenarios where heterogeneous integration (with disparate technology dice) will lead as opposed to homogeneous logic partitioned across multiple dice. There have been already memory products announced by Samsung and Micron (Hypercube).
For homogeneous logic partitioned across multiple dice, in addition to the advantage mentioned, there are another compelling reasons to do so:
-chip to chip interconnection where routing on the board and the number of interconnects makes it virtually impossible, leaving stacking in 3D or 2.5D as the only alternative. There are plenty of application scenarios in high speed networking.
-real estate space in XY for mobiles and handhelds is another reason where homogeneous logic may have to be partitioned across multiple substrates.