We dont know if there is any situation, it is simply even though both consoles are RDNA2 they are so very different they are not comparable in TF until we compare how things run.
CU do the shader work, and are given data by memory >>>> L2 cache >>>>>>>> L1 cache >>>>>>>>>> L0 Cache and you can think of each >>> as bandwidth, and the size of the cache and amount of compression is a factor in pumping data to the CUs.
Each step getting closer to the CU gets faster and shorter in distance and clock cycles. You cna see this in the RDNA1 white paper below, look at the bandwidths of the caches, RDNA1 is so much faster than GCN and even with RDNA1 having much less TF it smokes GCN.
How do Ps5 and XSX compare, well they both use 4 arrays, each fed by L1 cache. See below, we dont know the layout of ps5, but it will be similar concept with 4 shader arrays as well so I highlighted blue for easy comparison,.
So you cannot compare TF ONLY when one part has 20 % faster clock and the data paths will be very different, its not a simple one has 18 % more TF therefore thats the difference. The extreme example is RDNA1 vs VEGA
My take is XSX will have an edge, but not as much as TF analysis says, but we have to see software running as Cache size and efficiency is just as important as theoretical TF, as you have to keep CU utilised with work. Lets see eh ?