Tbar,
I don't have exact measurements on me, but there have been some changes which could justify the increase. Also your measuring method may be inaccurate
1, Rev 504 uses the new lighting, which eats up almost 8 MiB of RAM just being there. It needs that memory for the speed - it caches 3x3 chunks of data so that it can light an entire chunk really really fast.
2, Rev 504's generator uses a few buffers on its own. I don't have a number now, but I believe it could be a megabyte or two.
3, Chunk usage may be different for the versions - I'm not sure if 466 doesn't have a debugging mistake committed in where the viewdistance is set low by default. That would mean that each client sees less chunks and therefore uses less memory. Use the "/viewdistance <N>" command in the in-game chat to set the client's viewdistance.
4, MCServer doesn't unload chunks immediately. If you've used Rev 504 for a while, perhaps walked a bit, the number of loaded chunks (and therefore memory footprint) will be higher. You can issue "numchunks" command in the server console to see the number of chunks, and "save-all", then "unload" to minimize the number of chunks to only those visible to players. Those commands work for both revisions. Rev 504 also has a "chunkstats" command that you can use to see how many loaded chunks there are and how many of them are dirty (modified and unsaved).
cb88, I dont think MCServer could run reasonably in such an environment, tens of MHz just might not be enough. Dunno, I'd be glad if you proved me wrong, though
I'd say the 2x50 MHz setup would be better in this case, since the server is multithreaded and can balance its load somewhat. Although it depends on the OS's capability to run threads from a single process on two CPUs, I suppose it should be able to, but who knows...