Sephiroth; HSR wont kill the memory bandwidth problem that NVIDIA has been plagued by.
NV20 is a new design ground-up, it was talked about back in the nVidia interviews when original GeForce (NV10, just the codename was known) was just being released. Chip was paraller developed with GeForce, and has thus been in the works for two years. Rumours say same about the Rampage. Do you honestly believe that two ground-up nVidia designs would suffer from same, wasteful use of memory bandwidth? Hardware HSR is granted with NV20, I'm thinking memory BW effifiency is boosted further by also implementing larger vertex and texture caches.
NV20 is not GeForce2 GTS with HSR slapped into it. Thus it cannot be dismissed as yet another memory-starved product from nVidia at this stage.
Soccerman; BTW, if this chip has any of the features I have suspected for a while now, that means that it's specs are:
-800 Mpixels
-1600Mtexels
Aren't the rumoured specs indicating a four-pipeline four-texture unit rasterizer for NV20? As far as I can remember, DirectX8 specs call for four-layer texturing, and being able to do it in single-pass would certainly be a big advantage. That would mean 800Mpixs/s and 2400Mtextels/s peak fillrate.
But what really interests me is will those texturing units be able to do trilinear mip mapping in one pass, or will there be a 50% textel peak rate hit as is the case with GTS.