So how many of those "cores" or "nodes" can you add to the full GPU before you hit some nasty bottleneck? How much work will be duplicated in those nodes (triangle setup?)?
For the current generation of Series5XT cores one could theoretically connect up to 16 cores (MP16).
From the whitepaper Roderic posted above section 4.7:
The SGX-MP architecture is designed to scale as close to linearly as possible as more graphics cores are used. In typical real-world performance conditions (running well known game engines, graphics applications, benchmarks etc), each additional core runs at 95%+ the efficiency of a single core. Additionally, adding another core to the system only increases the overall memory bandwidth for a frame by <1%.
A SGX543 or 544MP16@200MHz would result on paper to Xenos XBox360 graphics performance.
Their new up to DX11.x generation Series6/Rogue, hasn't been fully announced yet and details over it are sparse. However from the two so far announced cores G6200 & G6400 the latter should be roughly on the fore mentioned MP16 level. Cores should be quite more efficient this time and probably scalable to higher amounts than with Series5XT.