AIDA64's benchmark is accurate and repeatable enough to compare similar systems. If someone is seeing a 10ns difference in latency, they're almost certainly seeing a difference in performance due to the same factors causing the differential in the benchmark. I plug the same timings into any given AM5 board with any given AM5 CPU of the same code name, run an OS with similar settings, and I'm going to get the same AIDA results, plus or minus some modest training variance.
Anyway, dual-CCX parts can take advantage of the bandwidth higher memory clocks can provide, but only if you're actually using both CCXes simultaneously in memory heavy tasks. For quite a few workloads, and most everything when only a single-CCX is actively being used, the 1:1 UCLK:MCLK ratios will be at least as fast.
I've got the 8200 version of that Team kit and can top out at about 8200MT/s with most Granite Ridge samples I've tried on a cheap two-DIMM (B650M-HDV/M.2) board, but 8000-81000 is faster due to the much tighter timings I can run, and 6200 is faster still on my 9700X. With a 9950X, I'd probably aim for 8100.
That said, if price/performance ratio matters, I'd still be looking at 7200 CL34 stuff, then either running 8000-8100 if the memory could handle it, or 6200-6400 1:1 otherwise.