I'm sure that there is at least some sort of trade off with respect to total package power and peak clocks due to thermals. While the L3 due aren't big power hogs, they are also certainly not free.
Like there was a trade-off with going to chiplets.
If you go from 1 to 2, it may seem like you are barely making up the lost ground. But you can go with 8 chiplets, and blow the doors off everything else on the market.
It seems like people are having hard time believing that the 3D stacking likewise opens a new dimension to performance scaling as the chiplet approach did.
AMD did not hold back with chiplets, went all in and gained performance crown in most important (server) market. Some people may have said that gong with 8 chiplets and 64 cores was an overkill, will use too much power, will have too much overhead there will be diminishing returns and all that. AMD went there.
Why are people thinking that AMD will / should hold back with L3 stacking? When, unlike adding chiplets, the incremental costs of adding extra layers of L3 are ridiculously low, the potential performance gains have fewest trade-offs / drawbacks?
If 4, 8, 12 layers of L3 delivers bigger performance gains on some applications than Zen 3 -> Zen 4 or even Zen 5, there is no reason not to go there....
Apple is certainly not going to hold back. Apple does not owe anything to the rest of the PC market.
Just because Intel and AMD held back performance gains from adding DRAM on CPU package, for benefit of OEMs, Apple trampled over this silly arrangement and went ahead with DRAM in CPU package...
If AMD holds back with going all the way in on SRAM stacking, there are too many competitors out there, and someone will use it fully and market opportunity will be lost.