I think the two big changes are that the nodes have 10Gb network cards and full bisection bandwidth. They're also HVM instead of PVM but I don't know how much that matters for this benchmark.
Internode bandwidth seemed to be the limiting factor cited in the paper, so that makes sense.
It looks like they are using extra-large standard and extra-large high CPU instances rather than the more recent quadruple extra-large "cluster compute" instances. Maybe that's the reason? (EDIT: Though amock seems to have said it better than me.)