The first benchmark chart of LLama.cpp is effectively a memory bandwidth chart, unsurprisingly. Each of the cards except the 5090 gets almost exactly 0.1 token/s per GB/s memory bandwidth.

               Mem GB/s  tok/s  tok/s pr GB/s
    RTX 5090    1792     158.77    0.089
    RTX 4090    1008     100.51    0.100
    RTX 3090     936      90.88    0.097
    RTX 4080S    736      77.08    0.105
    RTX 4080     717      74.56    0.104
    RTX 4070TiS  672      70.83    0.105
    RTX 4070S    504      54.59    0.108
    RTX 4070     504      54.47    0.108

The 5090 is an outlier in that it has 78% more memory bandwidth than the 4090, and "only" gets 0.09 token/s per GB/s. So it seems to start running into compute limitations.