1GB on a 256-bit is *NOT* in and of itself 'useless', I wish people would stop saying that, cause it ain't necessarily true.
To answer an above question, when you run SLI, your total memory capacity is whatever is available on ONE of your cards. This is because both cards have to have their own copies of all the same memory information so that they can alternate rendering frames without having to pass data between them over the PCI-Ex bus, which would be far too slow.
When you start talking SLI, especially with really high-end cores like the G92, you begin to have sufficient processing power to be able to run newer games at 1920x1200, with 4xAA and whatnot. This is *precisely* the scenario in which having extra memory begins to come in very handy.
If was thinking of going SLI with 8800/9800gtx, I would *absolutely* want two 1G cards. This means, I would not buy a GX2, obviously. With a single card, no, I'd probably stick with 512MB, because with one core you still don't have the processing power to run at high enough settings to *require* >512MB of vram (although there are exceptions to this as well, if you are talking a game that loads a lot of high-res textures, for example).
SLI would be a different story altogether.
This is especially true of the 9800GTX, given that we're starting to see some pretty dang high memory overclocks with this card, which will help with overcoming the limitations that might be posed by the 256-bit bus. Memory BANDWIDTH can come into play with regards to choosing a 'proper' memory capacity for a given core, but its ultimately still mostly about the core's power, and bus width is only ONE factor in determining total available memory bandwidth.