Quote Originally Posted by neliz View Post
Or a different way to access a pool of system memory, fast.

Even though memory throughput to the GPU is not a bottleneck, the limited amounts are. I'd like to see AMD and NV invest in a new memory transfer protocol so that graphics chips can enjoy fast transfers from their VRAM but have a faster way to access system RAM than todays normal Memory caching methods. Since this is evident with yesteryears architecture (GT200 chokes when it runs out of memory) I can only hope that NV (and ati) look further than wait for PCIe 3.0 where basically doubling bandwidth to 32GB/s only puts it around 15% of the VRAM bandwidth.

Solving that bottleneck would allow for massive amounts of textures in game.
Other way around. Bandwidth IS THE bottleneck. The 4k x 4k texture mentioned earlier is only 5-10MB depending on compression. You can easily fit dozens of these.

If you had no mipmaps, and the wall, door, ceiling, floor and table (5+ normal maps, + shadow maps) each used 4kx4k, even though each would occupy small amount of your 19x12 screen (and hence overkill, little benefit unless you stick your nose up against the wall, and why nobody does this), the bandwidth demand would bring even 5870 down to 1fps. Ofcourse, game developers dont make games just for people with 2GB cards.. they need to provide good gameplay on common platform.. ie 8800GT 512MB.

Its actually quite easy to test this in real life. Hookup a couple 30" LCDs. Fire up GoogleEarth. Monitor memory usage with Rivatuner. Maximize GoogleEarth across both screens (wait for it to load and cache everything).. you now have equivalent of a 8MP texture.

How is the frame rate compared to the memory usage...?