Yes that's what I also read after posting this. The 3 fold increase is over dual Xeon I believe. So that's still almost 6 times over a single Xeon.
It seems that in some cases the benefit can be much bigger, like in financial monte carlo modelling. I don't know if it's remotely connected to Monte Carlo GI
but what it does show is that it's quite dependant on the task.
8GB may seem low, but that's still roughly 150MB per core/thread. TG2, for instance, can use 50MB as lowest amount of RAM allocated to each thread. The RAM is only for execution, not for storage.
Scene data would be stored into the system/node's RAM, which would obviously be bigger.