That's not what I got from the quote.
Until we get the actual details, what I said could be the case.
We actually have all the details. The VGLeaks leak has been verified to be spot on by several reliable independent parties, including Digital Foundry. Also, using such separated system would be completely pointless, given how the whole point in Jaguar, the thing AMD advertised about it for potential customers, was how it can be quickly integrated into a semi-custom solution with whatever else the customer wants.
I would like to add. AMD doesn't seem to produce any 12 or 18 GCN layouts.
There has to be some reason for such modularity.
The only layouts AMD is producing are 32, 20, 14, 10 and 6. The rest are harvests. In one of the technical presentations near GCN release the AMD rep said that they can integrate the CUs in any amounts they want. Only, masks are expensive, so you only want enough chips to serve all market segments, and since vendors don't want too many models with very similar performance, you limit harvesting to about 3 models per chip.
The consoles won't physically have 12 and 18 CUs. Redundancy and harvesting is absolutely necessary for yields when building large chips on the TSMC 28nm process. By the last numbers they released (somewhat oudated now) their pre-harvesting yields were ~40% for a much smaller chip. Fully enabled chips can *only* be produced economically if you can harvest the failed ones into lower-value models. When you can't do that, you must include some redundancy into all chips so that you can use the majority of chips that have a failure or two.
Given the programmer-visible CU counts, I'd expect there to actually be 14 and 20 CUs on the chip.