I'm running one 500GB database on a 64-bit, 8GB VMware virtual machine, with 2 vcores, PG 8.3.9 with shared_buffers set to 2GB, and it works great. However, it's a modest workload, most of the database is archival for data mining, and the "working set" for routine OLTP is pretty modest and easily fits in the 2GB, and it's back-ended on to a pretty decent EMC Clariion FibreChannel array. Not the typical case.
For physical x86 servers, brand name (e.g. Kingston) ECC memory is down to $25 per GB in 4GB DIMMs, and $36 per GB in 8GB DIMMs .... dollars to doughnuts you have a server somewhere with 2GB or 4GB parts that can be pulled and replaced with double the density, et voila, an extra 16GB of RAM for about $500.
Lots and lots of RAM is absolutely, positively a no-brainer when trying to make a DB go fast. If for no other reason than people get all starry eyed at GHz numbers, almost all computers tend to be CPU heavy and RAM light in their factory configs. I build a new little server for the house every 3-5 years, using desktop parts, and give it a mid-life upgrade with bigger drives and doubling the RAM density.
Big banks running huge Oracle OLTP setups use the strategy of essentially keeping the whole thing in RAM .... HP shifts a lot of Superdome's maxed out with 2TB of RAM into this market - and that RAM costs a lot more than $25 a gig ;-)
