I find your memory bandwidth limitation theory interesting. Have you seen any supporting evidence elsewhere, beyond your own benchmarks?
Its a long video but you should really
watch all of it, its where my theory comes from.
7900 GRE: 80 CU's @ 2.4 Ghz = memory bandwidth <100%>
6950 XT 80 CU's @ 2.4 Ghz = memory bandwidth <100%>
7800 XT 60 CU's @ 2.4 Ghz = memory bandwidth <120%>
6800 60 CU's @ 2.4 Ghz = memory bandwidth <100%>
The 7800 XT is underclocked on the core.
The 6800 is overclocked on both the core and memory.
Add to that the 7900 GRE with 80 CU's and 18 Gb/s memory is the same performance as the 7800 XT with 60 CU's and 19.5 Gb/s memory.
The IPC RDNA 3 vs RDNA 2 is quite a lot higher, but very clearly completely hamstrung by memory bandwidth, well, the only one that isn't is the 7800 XT, which is why for a small 60 CU's GPU its the one that seems unusually fast in the range and is also the one that overclock's like a beast.