If the card has disabled units, as I expect it will, we can potentially expect to see 1.5GB of the 12GB that operates like the 500MB crippled VRAM on the 970. I'd expect nvidia not to bother concealing that this time though and it not to have much effect on performance. 12GB is beyond sanity of what you need for gaming. 6GB would be ample for any use case. Maybe three 4K monitors in surround would push beyond that.
I highly, highly doubt they'll do that this time around. I also don't expect there to be disabled SMMs.
If they do have disabled SMMs, they do not have to disable ROPs and the accompanied block of L2 cache.
Or if they feel it is so necessary to disable a ROP/L2 block, then they should not introduce a crosslink with the neighboring ROP/L2/MC chain in the shared partition. This cuts out an entire segment of VRAM, but it entirely slashes the potential for a forced period of low-bandwidth memory throughput.
They either have die issues during production that causes some disabled ROP/L2 blocks, or they intentionally disabled that. If it simply happens and they need to use the chips, they should preferably go a step further and simply prevent unnecessary performance issues with a partitioned small block of VRAM by disabling access to that one block of VRAM entirely. Or if they want more round numbers or whatever, disable an entire partition. Then the final card has 5GB (or 10GB) of full speed memory, and no slow partition to cause performance degradation.
I don't think the Titan X will have any disabled ROPs. If they are having chip issues that may force them to have disabled ROPs in another product, I think they'll release a cut-down 980 Ti or a 985/990, and save a fully-functioning GM200 to one-up that some time later. And I think the best strategy to save face is again, disable an entire ROP/L2/MC partition and release a 5GB card, with 2x512MB blocks of VRAM completely inaccessible.
Having a max VRAM limit is better than having a soft-limit that, once crossed, degrades all performance. A max VRAM limit won't have such performance degradation.
If the full GM200 is 24 SMMs, then perhaps a cut-down model has 20 SMMs, in which case, dropping an ROP/L2/MC partition would bring down max performance by 16 pixels/clock, to 80 from 96. If it has 20 SMMs, that is the absolute limit it could produce anyway, so the extra is unnecessary. Plus, that part will STILL have superior performance to the 980, and would carry a 320bit memory interface. They could even later release a 10GB model if they so chose, and in an effort to cut costs, could produce PCBs that actually only had 10 RAM chips instead of 12. Something tells me they wouldn't actually go that route though, so that the PCB could be utilized for a full GM200 as well.