I know people dismiss TFLOPs a nonsense comparison, but when comparing cards from the same manufacturer and using the same architecture it will be a pretty good indicator of performance. The TFLOP number includes both frequency and CU count. As far as I'm concerned no evidence yet found has convinced me that because some GPU tasks favour frequency over CU count the 20% TFLOP difference will be overcome.
Digital foundry touch on it
here (~11-14 mins) but with RDNA1 cards comparing 5700 (36 CU, 2100 MHz) vs 5700 XT (40 CU, 1890 MHz). So both at identical 9.67 TFLOPs. The 5700 XT still comes out with an average 2.5% performance advantage.
There are still some uncertainties. RDNA2 might have different performance scaling for different tasks, the PS5/XSX CU count difference is much larger (16 CUs) and the frequency difference is also larger than the above example (20%).
So basically previous evidence points to XSX GPU outperforming PS5 GPU, unless something new comes to light I don't see a reason to believe otherwise.