More I look at your score, more convinced I am, that your 32,5ns latency test is bugged. I can produce similar / better scores.
If you bench same settings like 10x one or two of those will give you bugged results, which look almost normal, but are not real.

Looka at this new screen, where I tweaked subtimings a lot and got nice direct scaling from my previous score. I can reproduce this as many times I want.
Also aida is scaling directly with the memory speed and I have not seen any bugs in it even it is runned multiple times in a row?



So, I think that is the highest real score so far.