Memory bandwidth of 3080 is only 1.7x of 2080 so that may bottleneck rather than TF. We haven't seen actual RDNA2 perf with the predicted CU/memory configurations so I don't see how you can compare perf/TF.
its really simple, they haven't done anything architecturally that we've seen revealed thus far that indicates any regression in perf/TF vs rdna1 here, and infact we know of several new features that rdna2 supports that rdna1 doesn't that will mean more perf/tf. But even doing this math assuming no gains whatsoever, we can say for a fact that perf/TF is better on rdna2 than on ampere for gaming. (rdna1 perf/tf is roughly the same as turing perf/tf, meanwhile ampere perf/tf is much lower than turing perf/tf and rdna1 perf/tf), so there's your point of comparison for this napkin math estimation. Basically, with everything known about rdna2 its impossible for perf/tf to have regressed at all, let alone regressed as much as ampere did vs turing. Its not really a useful metric comparing different architectures in either case, its just funny that perf/TF used to be a big nvidia talking point in these kinds of discussions and now it suddenly doesnt matter at all, just like power draw no longer matters at all and it used to be the most important thing in the world.