Initially I had a similar idea of using some sort of daughter board,
desolder original memory chips and solder thin flex wires to the mounting points of the chips. The problem with that setup was that I needed a lot of through holes on the daughter board since I cannot make routing lines between the 0.8mm pitch of the chip legs (at least not in my home-brewed prototyping lab), so I gave in to the stacking solution.

You see, I'm not really satisfied with the memory size until my hard drives spin only once per each play list (e.g. an album).

Besides, the SA1100 chip does not have math co-processor, so that makes it slower than say Pentium 200 Mhz.