That's interesting Serinity and certainly for me you're leaning on an open door because I was an AMD fan and all my builds were AMD without fail. Then they lost their way and Intel was/still is the only decent option.
AMD never really "lost their way", they were beaten down by Intel's unethical and illegal business practices. This isn't me being a fanboy before someone says it either, Intel's offices in several countries were even raided by police. Intel was found guilty on several counts, but the damage was already done and Intel had already gained so much out of it. They put us in about a decade of slow innovation because AMD didn't have the funds to compete anymore (remember how I said it takes 8 years to create a new CPU architecture? Nice coincidence eh?).
Zen was created with AMD's then horrible budget in mind. The CCXes (each die) are all the same 'template' for silicon fabrication. Each one of these template plates costs nearly a billion to produce. They legitimately did not have enough money for more than one die type. That's why AMD created infinity fabric, the lanes in the organic substrate that connect the CCXes. The highest end EPYC chip and lowest end Athlon chip use the same template (how neat is that?).
So by Intel being so violent and putting AMD in such a bad place, they ironically put AMD in a position to where they had to create an architecture so revolutionary and so cheap to produce that Intel will literally not be able to keep up.
Now, there are limits to the current implimentation of Ryzen. One of the biggest limitations being the lanes through the organic substrate, and another being the memory access. The memory speed also dictates the speed that infinity fabric runs at, which is why Ryzen is much more sensitive to memory speeds. AMD knew these limitations but didn't have the funds to correct them, but now they do. Enter Zen 2. The memory controller / IO / etc. are all moved onto a separate die, with that die being surrounded by smaller dies that contain nothing but cores. This makes the core dies smaller, significantly increasing yields even further. The memory controller and such aren't as sensitive to process size, so they're likely on 22nm. This new approach allows for more dedicated space to a larger and more robust memory controller, a significant amount of more cache (there's a rumored 400MB (!!!) cache EPYC processor), etc. This will need special routing and possibly an active interposer, one with logic built in.
(pssst, look at the authors of the papers).
An active interposer with a reverse butter-donut method of routing wouldn't cost them much more at all (potentially even cheaper) at higher core counts and would allow for significantly reduced latency as well (another problem that infinity fabric introduced).