GDDR5 is currently the newest GDDR available and is actually faster than the desktop DDR3 in almost all cases.
Depending on the stick (for a desktop pc), it can have 16 (8per side) 128MB modules, or 8 (4 per side) 256MB modules on a 2GB stick. 256x16 for 4GB sticks (though i had heard you can get 512MB modules now).
High Density memory, whether it be DDR or GDDR costs more, and the prices do not scale linearly, the reason for this is high density ram is harder to manufacture at high yield. The GDDR5 on the GTX 580 consists of 12 x 128MB modules, totalling 1536MB's of ram. Given the pricing of 4GB sticks based on 256MB modules.... the majority of the boards cost to manufacture would end up being the memory itself and would usually be sold at a loss. .... i expect the 3GB 580's are in being sold at loss and being paid for by the lower end 400 and 500 cards hot cake sales.....
Another aspect is that Current DDR designs are still based on 64bit channels, which is the reason why even DDR5 at the same clocks will perform slower than on a Video card, which uses 12 x 32bit channels connected to 6 x 64bit memory partitions in the GPU.
The operating system does provide a spilloff ala "shared video/texture memory" in the main system memory, but i don't recommend bumping everything to high and relying on this shared memory because then i would have to hunt you down and shoot your computer for posting a "Why is my game so slow" thread.
If one could put the GDDR5 from a 580 onto a motherboard, specifically an i7 triple channel setup, when spread across those 3 channels it would be 96.2GB/s, on a i7 2 setup it would be 64GB/s.
There is also another factor, which is latency, and not just the latency of the modules, but also of having to traverse the system bus, and be directed by the cpu to the memory. ..... i won't even attempt to imagine the input lag this would incur when applying AA.....
post edited by squall-leonhart - 2011/03/29 16:17:52