- Taking a bet? I'd suspect no, otherwise they'd have done it a long time ago since this really isn't anything new, its just throwing more of the same at the problem. It will undoubtedly help, but the trade off will end up being absolutely gargantuan die sizes (and prices, which NV now knows people are willing to pay).
AMD has some outstanding EEs and hardware devs and RDNA2 looks like it was the result of some smart collaboration between their CPU team and their GPU team. NV certainly has some brilliant engineers no doubt, but they've generally come off as a much more iterative in their designs and less likely to swing for the fences like AMD is prone to doing
Eh, I'd say they're definitely going down the LLC route, same as intel will. Bandwidth optimization is a problem across the entire high performance industry, and if one option works well for one company it'll likely work at least decently for every other. Thus a concentration on sram cache, same as with CPUs.
In fact if you really want to guess at the specs of next gen cards you can just look at bandwidth limitations. Worked nigh perfectly for me with RDNA2 and Ampere, and I'm willing to bet it'll be similar again this year. 512bit buses are in the seeming "red zone" for chip makers, who far prefer going out of their way to acquire faster ram over building a bus that big for whatever reason. HBM is an option but significantly more expensive than GDDR, thus at the farthest end might end up on $2000+ cards. That leaves us with faster ram, and the fastest possible ram this year is 24gbps announced by Samsung, but the timeline for that was vague at best. Maybe 20gbps might be much more commonly available, and 21gbps for GDDRX.
That leaves giant caches as basically the only option for Nvidia. We can't assume their engineers are magical all seeing ones anymore than the ones that work at AMD or elsewhere, so one was gotten over on them a bit. Though I can see the logic behind preferring faster access to main ram. As RDNA2 demonstrates, a giant cache isn't some universal silver bullet, especially against deferred rendering which needs a ton of bandwidth to fill it's gbuffer and where RDNA2 obviously suffers performance issues.
Which is also why I don't believe the rumors of RDNA3 maxing out at a 256bit bus for even a moment. The 6900xt is already performance capped in numerous titles by it's limited ram bandwidth versus the 6800xt, which it should be over 20% more performant than across the board but isn't, thus the refresh with 18gbps ram that's coming up. Thus we can see that in all likelihood Nvidia will get a bigger cache, AMD will get a bigger bus, and we'll just go from there.