There's also a matter of personal expectations. Some folks are perfectly fine with operations taking 20 seconds, others find operations that take longer than 2 seconds to be intolerable. I recommend working at a medium resolution (4kx2k or so) to get the world and then do more local areas as has been suggested. This technique will balance the amount of data the machine has to handle against the quality of the final output. And, the bigger the machine, the larger the map before you hit the slowdown where it has to swap to disk continually.
The biggest problem with this technique is that certain whole-image operations are extremely sensitive to the size of a pixel in order to get plausible results. The most common example is fluvial erosion. In the case of erosion, a river is almost never even a pixel wide, but most flow analysis and erosion algorithms require and/or generate rivers more than one pixel wide. The results are beautiful, but they have a certain signature to them, just as most simple fractally-generated terrains do.