Most (if not all) todays graphic cards do not contain any cache…from what I can remember.
Umm, that's not remotely true. Today's cards have plenty of cache. Texture cache is there, but if you write your pixels and access your data intelligently you don't need a whole lot. More on-chip memory is found in fifos/register cache/pixel cache/whatever you want to call it, and there's a ton in any graphics chip. This is what makes them good at absorbing texture latency and sustaining their peak throughput. You also have caches for storing information about Z and colour compression, and in the case of ATI, for Hi-Z data.
RSX having cache guarantees to a certain existent faster texture swapping, less redundant data processing, faster post processing routines, ECT…
That's pretty much all wrong. Only in the worst cases (e.g. tons of textures, or very non-coherent dependent texture access) is texture swapping an issue, in which case more cache isn't the solution. Post processing has very predictable memory access, and again more cache won't help much. Don't know where you could save redundant data processing.
I guarantee you that G70 has already optimized the cost/benefit ratio of caches to the point where you'll only get a few percent more performance with bigger caches. Any tricks to maximize memory usage are already in use now to minimize costs and maximize profits for the low bandwidth budget cards in the PC sector.
All cache does is help you achieve perfect efficiency. For bilinear texture filtering you need the nearest 4 texels, and under magnification many neighbouring pixels will need the same 4 texels, and that's what the cache is for. Even under minification, adjacent pixels need 2 of those texels. If you draw your pixels in an intelligent order (and believe me, these chips have done so since the GF3/Radeon days), you're unlikely to need to load many of those texels again.
There's a fundamental limit to how little bandwidth you'll need, and modern GPU's from ATI and NVidia aren't far from achieving that.
The only time cache will actually save you significant bandwidth is if a whole texture can fit in the cache and that texture is tiled and repeated, which we'll see less of in
next-gen graphics, or if you're doing tiled rendering like Xenos or PVR's TBDR chips do.