I can only conclude that you jumped straight into this thread without reading previous posts by several individuals.Yawn, 2016 and the multi-GPU small-die flagship theorists are back in full force, AFTER AMD has commented that they want to move away from budget image and AFTER AMD delayed Fury X2. Considering how we have seen some major AAA games not have SLI/CF support, how no UE4 games support SLI/CF (as far as I am aware), and considering how both AMD/NV just unleashed 596-601mm2 28nm flagship die products for $650, the trends cannot be clearer. What's going to happen is as the 14nm/16nm node matures, AMD/NV will repeat the 28nm node by slowly moving from 275-350mm2 die towards a 600mm2 die over the next 3-5 years. Then, the strategy will restart with 7/10nm node. It may be true that along the way we could see a dual-chip card with mid-range die aka GTX690/7990 style but there is no logical reason to suggest that AMD/NV will stop pushing the die size limits along the node maturity curve.
Perhaps, most important of all is the consumer's buying trends -- to this date, no dual-chip flagship card from ATi/AMD/NV has ever sold well. Even the OG Titan sold 4X quicker than GTX690 according to NV (http://www.pcgamer.com/nvidias-surp...e-year-old-gtx-690-in-just-3-months-were-not/). Even when R9 295X2 cost almost the same as the much slower 980, almost no one cared for the cooler, quieter and way, way faster R9 295X2. Even if the technology and cost allow for dual mid-range + interposer concoction, the consumers have long voted that 1 large monolithic die, even with way less performance is more preferable to them. It would be a business mistake for both AMD/NV to not see these trends, along with a buckets of data going back to HD4870X2/GTX295 that just underlines what I typed above.
It seems too many can't see past SLI/XFire when they read multi die composite GPU. Discard that thinking.
There are several possible solutions for having a composite GPU function and appear as a monolithic one. One is a central core die with all IP blocks except [geometry, CU, rasterizer and rops]. Latter block is replicated as needed. From one of the research papers linked earlier "the impedance across the interposer is identical to conventional on-chip interconnects".
When you look at a die shot of a GPU, don't you see blocks that have specific functions? How are these connected on a monolithic GPU? Don't we use routing through the die layers? All an interposer will do is substitute the already existing on-die connections in a monolithic design for an-off die interposer connection between the separated IP blocks.
Also don't get stuck thinking that the multi die must be composed of identical sub-units. It can be asymmetrical and still have the ability to scale.
You acknowledge that AMD wants to increase ASP and change their image as the low price option. This strategy is the ideal way to achieve this goal.
What if they can offer a big die [600mm^2] equivalent GPU this year? What do you think their ability to price will be?
You see a gradual approach to the 600mm^2 monolithic GPU as a given.
The interposer can be larger than the Fiji sized one. We're acting as if the interposer has active elements. It's a piece of silicon with copper routing. The design problem is an interconnect layout that is able to be done with smaller reticle limited exposures. This was done with Fiji, so they have experience.
In this case, why do you think that 600mm^2 is the limit for a single GPU?
What if they can offer a bigger than 600mm^2 GPU next year? What do you think their ability to price will be?
One of the papers linked had a cost analysis for what looked to be a mature process, [91.7% yield @ 148.5mm^2]. It is cheaper to do multi die than monolithic as you already have an interposer. I think it's safe to assume a greater benefit with a lower yielding new process.
The ability to obtain a greater bin range is icing on the cake.