If those 625M transistors are also carrying the dedicated DSP, (rumoured) Cortex-A level ARM CPU for I/O and "sleep activity", embedded DRAM and other stuff, then it's really disappointing for what's left for the GPU itself...
Why on gods' green earth would you assume all that? (Cortex A5 isn't terribly heavy btw, but even an ARM11 is overkill for what you outline, and requires a grand total of 1mm2 on 40nm.)
I distinctly remember the fully custom PS3 GPU, as both Kutaragi and Jen-Hsun assured us it was. Turned out to be an almost plain vanilla product, with some CPU-GPU communications bolted on. That will be needed on a Nintendo GPU as well, on the other hand there is limited need for PCI-express, and even less for IEEE754 rounding compliance...
We don't know that the implied part is a Nintendo product, nor do we know if the part is intended to be used exactly as described in the case study or if specific functional blocks are readily modified, and a plethora of other things. However,
if we assume that the described part is indeed the Nintendo GPU, then we can safely dismiss GPU e-dram for instance. Why would a GPU of those dimensions, on a GDDR5 memory subsystem, need e-dram? The CPU will have it, we know that from the horses mouth, IBM. But the GPU?
If they are integrating stuff onto the GPU, it is likely the above mentioned CPU<->GPU bus, plus the usual housekeeping. Nothing that takes a lot of die area. It just isn't needed assuming a dedicated die for the CPU, and GDDR5 memory. (I'm assuming that whatever is needed for over-the-air communication with the tablet controller resides on its own chip.)
It is seductively easy to assume that wsippels lead is indeed Nintendos upcoming GPU, but it seems very premature to take it for granted. I sincerely doubt we'll ever know for sure. Even a year after the release of the 3DS, the real specs for its GPU still isn't public knowledge.