Read the article I linked to in my previous post, one of the better explanations out there. Basically it's memory addresses reserved for input/out devices. It dates back to the earliest systems, and has been kept for backwards compatibility. The largest chunk of the MMIO (memory mapped input output) is taken by video memory. Basically a 32-bit operating system can only work with roughly 4GB of total memory. So if your GPU has 512MB of video RAM, Windows will work with that 512MB and only has around 3.5GB left for system RAM. Other devices have memory addresses specifically reserved for their use, so when you add it all up, only about 3.2GB of system memory can be used. If your GPU is a 1GB card, or worse, you have several video cards each using 512MB or more of video memory, the available system memory could be even less. 64-bit OS can address beyond 4GB of memory, so multi GPU systems will benefit more from 64-bit OS. But the MMIO hole between 3.2 and 4.0GB still exists for both 32 and 64-bit unless you enable memory hole remapping.
But as the article points out, installing more than 4GB of memory is just the easiest way to combat the problem (though not the cheapest). But quality 2x2GB 1066 and 800 kits can be had for $30-55 shipped, so it really isn't a big deal IMO.