Of course, you'd probably be best to hope that you don't have to use system memory instead of the videocard's own RAM. I'll elaborate on SpideyCU's example of the hard drive serving as virtual RAM: RAM, being solid state is thousands of times faster than the hard drive. RAM accesses data in nanoseconds; HD's do it in milliseconds. RAM deals in GB/sec. HD's deal in MB/sec. So anyway, here, virtual memory is slow.
It's the same way with a videocard; it can use its own local RAM at high speed, but if it runs out there, it has to go over the system bus, and access the system RAM. Not only does that have an effect on latencies, but it also has do endure those latencies at the slower speed of the chipset and RAM. This will really kill framerates - it would quickly go from there being SO much data in the videocard RAM readily available to the video processor that it would slow down, to a situation where there's lots of data present, but the video processor can't get it from the system RAM fast enough to put out a good framerate.
In short, sufficient local video memory is a good thing.
Also, in case you're wondering what the point of this sharing is, it's in part because back when AGP was made available, high-speed RAM was pretty expensive. So, why not just make the big pool of system RAM easily available to the videocard? But RAM prices came down quick, and videocards pushed ahead, to speeds much faster than an AGP would be able to handle. AGP 8x is spec'd to run at 2.1GB/sec. New videocards can handle over 20GB/sec using the RAM on the card. Going over the system bus to the main RAM would utterly kill performance.
The AGP does still do an ok job; it provides a dedicated Port, not a bus, for the videocard to do its thing. But its days are numbered; PCI Express is coming.
Peter, feel free to check this through for technical accuracy. I wrote it out of my experience and understanding of current technology; yours seems to be far greater though.
