Also, even 32MB of embedded memory cannot compensate for lack of features. A GS32 could never produce the kind of visuals that a GF FX can - it can only push lots of cartoon-like polygons.
That GScube, did render scenes from FF movie and Antz nicely though, even though GS32 lacks features. I think you can substitute most features, by multi passing. You either go one way or the other I guess.
I think for the next iteration of GS, they would need to increase color precision, to match GF FX visuals.
Also 1 billion transistors chip, isn't too hard if most of them is spend on eDRAM. I am still wondering why, GPU maker for PC hasn't gone the eDRAM route. They could make a low end DX7 solution that's really fast using eDRAM. Considering PC developers are targeting low end machine anyway, features aren't that important, but speed are.