IF was designed as a powerful general purpose interconnect that can tie together different types of modules that might not even "know" how to talk to each other. It certainly isn't limited to connecting Zen cores together.
On the other hand for a gaming MCM GPU you need a more specialised interconnect - though there is nothing stopping them evolving IF to support that in future.
You don't really need a more specialised interconnect at all, what you need is a more specialised utilisation of infinity fabric. For all intents and purposes it's just a cache coherent protocol that can be pushed over any connection. You can run infinity fabric through pci-e, on an interposer, silicon or organic, the difference is how many lanes, how big the controller based on how much data it's pushing, how much power is used and thus the performance. To make multiple chips work together as one on the gpu side is just significantly harder than on the CPU because with CPUs you have very limited information that really wants to go to a single place with few options and it's damn easy to schedule, with a gpu half the battle is utilising the cores effectively with constantly changing tasks and differing workloads. I actually think they are underplaying it right now and it's definitely for the future.
But where a CPU can use 4 identical Zen dies for a gpu we almost certainly won't see that. What we'd most likely see is. Lets take a 700mm^2 die for comparison, instead of one monolithic die we'd see a 40-50mm^2 control block sitting in the middle, then you'd have say 4-6 blocks of shaders, say 512-1024 shaders in each block and those are 150mm^2 per piece, then you probably start dedicating things by making a separate ray tracing block and put that in, with a different number of connections than the shader blocks because at this point you can say ray tracing uses less or more bandwidth (whichever it is) so we'll optimise by having 30GB/s to each shader block but 15Gb/s to the ray tracing block.
If you had a single identical die then apart from having the same control parts of the die duplicated in each and wasting space, because the major battle in GPU is effective utilisation then everything is waiting on the master block to decide things and then you'll get stalls everywhere and one overworked main block that isn't designed to control 4 dies.
I don't think IF would have any trouble being used for such a layout and connections.
However IF being used on a motherboard is no different to pci-e, it won't enable some magical ability for chips to work on different cards to somehow behave as one. An IF connection routed over PCI-e bus ain't doing anything particularly new. NVlink can't enable magic modes, neither can XMI or whatever it's called, neither can using IF over a pci-e bus. Until you get a bus that allows the same bandwidth and latency that is available for on die communication, or at least in the same magnitude, then you aren't getting close to working as one die across slots.
