AMD Whitepaper is inferring it has architecture that increases teraflop performance with less basis on increased clock speed's/CU Counts and bandwidth - this is a welcomed/expected side effect / benefit of die shrinks and will only grow larger and larger the closer we get to 3 nanometer's and ultimately 1.3 nanometers and below.
The Teraflop will remain firm at the standard 4 trillion calculation's per second barring improvement's to the architecture meant solely to boost the calculation throughput - as is to be expected.
Architecture improvement's can also bolster a single teraflop's performance.
Hardware revisions/improvements will continue to see the teraflop calculation throughput bolstered even higher - past the standard 4 trillion calculation per second threshold it is at currently with mere software optimization and efficiency issues sorted out - I would say through hardware optimization's alone you may be able to squeeze out 4.7 trillion calculations a second bare minimum, but up to 10 Trillion calculation's a second and far more than that in the future - utilizing die shrink's as they are currently - but that will improve vastly and these same hardware improvements will permit those previously mentioned software improvement's to produce far.... far... greater result's.
Teraflop performance matter's less (not really - it always ALWAYS matters) when you take a machine learning program *COUGH*DLSS 5.0 *COUGH* that can recreate a scene in game verbatim utilizing very little data - this is what we refer to as disruptive performance gains. And plenty... plenty of those are on the horizon - along with pure software optimization that bolster's performance 100,000%. Lot's to be excited about in that Arena.
If you're a scientist looking for high performance compute from your GPUs, you're going to buy Quadro/Tesla or Radeon Instinct. In pure compute workloads more teraflops undoubtedly matter.
Games however, are not pure compute workloads. More FLOPS certainly is better, but what matters more is shader occupation and scaling. The more shaders you shove into a GPU, without proper scheduling (Be it via a hardware scheduler or driver scheduler) most of these shaders sit idle and do absolutely fucking nothing for your games.
See: Vega 64 Vs the 1080Ti. The Vega 64 had 4096 shaders and the 1080Ti had 3584. And yet the 1080Ti butchered the Vega 64 in games by an average of 30%.
See also: Ampere Vs RDNA2. The 3080 has 8704 shaders where the 6800XT has a paltry 4608. And yet the 6800XT is ±5% of the 3080 in gaming workloads. We're talking a ~30TF GPU VS a ~20TF GPU both performing near enough identically in games.
The only times the massive shader advantage plays to the 3080s benefit is at higher resolution where the massive number of pixels can fill up all those otherwise idle shaders.
Note also that in both examples above the faster clocking card had higher performance per FLOP in games:
Vega 64 @ 1500MHz on average vs. the 1080Ti @ 1900MHz on average.
3080 @ 1800MHz on average Vs the 6800XT @ 2300MHz average.
Of course the differences in architecture play a bigger role, but it's something to consider regardless.
Memory bandwidth helps improve performance per FLOP as well. Radeon VII has 3840 shaders but thanks to its massive 1024MB/s memory bandwidth it smokes Vega 64 with it's paltry 484MB/s at ISO clockspeed. It's the reason Nvidia spent so much time and money making the power hungry GDDR6X to feed it's massively compute heavy Ampere architecture.
Ultimately. Compute performance is compute performance. And that's great if you're a scientist who wants to do a lot of mathematics very fast. But this is a gaming forum and we're dealing with games. Games do not scale linearly with compute performance in the same way mathematics does (I mean if you look at actual compute benchmarks you'll see that even pure compute workloads often don't reach a given GPUs peak theoretical compute performance).
It's unwise to use compute performance as a barometer between two different GPU architectures. The PS5 and Xbox are theoretically the same GPU architectures so you should be able to compare them reasonably well. The Series X will undoubtedly have an advantage in compute bound games. But as we know games aren't purely compute bound. Th PS5 and Xbox both have 64 ROPs but the PS5 is clocked at 2230MHz Vs 1825MHz. So the PS5 will have a higher pixel fillrate. We have some inkling that the PS5 also has some unique geometry features that place it somewhat ahead of the series X. Even if both geometry pipelines for the pair of consoles were identical, both have 4 shader Engines so the PS5s higher clocks will give it a higher triangle throughput. I'm sure the Xbox also has some customisation that give it a feature advantage. But the ultimate point is that they're not both identical GPUs so things aren't quite so clear cut. There's also the matter of the CPUs. The Xbox has a higher clocked CPU, but the PS5 is reported to have a unified L3 cache. We know games are highly latency sensitive so the unified L3 cache will give the PS5 some advantage in gaming workloads in terms of CPU performance.
TL;DR. It's not that simple.