We did find GF117 hidden inside a driver, didn't we? As I said, GF117 being a 28nm GF116 but with a smaller 128-bit bus (with support for faster GDDR5) seems to make sense. Remember the main advantage of 28nm in the short-term is power rather than cost (lower yields & higher wafer costs) plus capacity is limited - so it makes sense to create a GPU that's perfectly suited to the OEM notebook market first and try to get some Ivy Bridge design wins with it. Of course, I could be wrong, or even if I'm right that doesn't mean NVIDIA will manage to deliver it in time.
Yes, they mention 10-20 PF peak as system goal. The 10 PF figure probably applies if they put only Fermi-Teslas (GF100 based 2050 or 2070) in.
It think they talk about the second PCI-Express-Slot. The GPU is connected to the GPU via PCI-E 2.0 according to the pdf. No idea what is supposed to be in the first slot, as the whole interconnect between the nodes goes over a HT connection.
Can't see why it would even need faster gddr5 memory. It's not terribly bandwidth limited with 128 bit gddr5 today, mobile parts are lower clocked and use (probably) only 1.35V gddr5.
Ah yes, good point about mobile parts being clocked lower for both core and memory. Obviously I'd expect any Fermi derivative on 28nm to be able to clock higher than GF116 on 40nm though, so you'd ideally still want to support faster GDDR5 for the desktop SKUs. That doesn't mean they'll bother though.
