There’s no single element that’s more important, you want to balance everything.
Regarding RAM, I just did a quick test to give you some concrete numbers. Starting from the default cube, I applied subsurf at level 3 to give a base mesh of 384 faces. (Not declaring this to be optimal or anything, just a test.) Then I added multires and switched into sculpt mode (sculpt mode draws differently than object mode, so subdividing can be faster in sculpt mode purely by virtue of not uploading as much VBO data to the graphics card.)
Subdividing up to multires level 7 (about 6.2 million faces) uses 1085 MB of memory (according to the info in the header.) After subdividing again (25.1 million faces), mem usage is 4193 MB. Sculpting is still usably smooth at this resolution. Note that this is RAM usage; the graphics memory used should be much lower.
So, taking 25 million polys as an example, this would probably be totally unusable with only 4GB of RAM. It would have gone into swap space during subdivision, and as soon as you’re in swap you’re dead. (Swap on an SSD helps some, but still probably quite slow.) On the other hand, 4GB of RAM at the 6 million poly level is probably plenty, and adding more RAM wouldn’t make 6 million-poly sculpting go any faster.
It’s also quite possible to run out of graphics memory, and this could cause all kinds of nastiness. Depending on your graphics drivers it might crash, revert to a slow path, or start swapping data between system RAM and the card. The “fancy” features of the graphics card (vertex/fragment/geometry shaders for example) won’t matter much for sculpting, since all the actual sculpt calculation stuff is done by the CPU.
As far as the choice of CPU, I don’t have a lot of solid data on what’s important. Faster is generally better, but if your real bottleneck is e.g. sending data to the graphics card, a faster processor won’t help any.
My guess is that you’d do well to upgrade to at least 6 or 8GB of RAM (quite cheap these days), but I’m not familiar enough with your processor and graphics card to offer recommendations on those.