I think by now it makes sense to assume there's some sort of malignant design in the Ada generation's silicon, that the mobile parts have a performance cap, or an intentional bottleneck of some sort. Otherwise I can't think of any reason why the cards hardly improve at all from ~105W to 140W yet when put on a discrete card and given ~300W can improve another ~50% from that ~105W performance level. Of course the efficiency curve isn't linear but a plateau cannot reasonably be assumed to exist at 105-140W while somehow sparing a large part of 140-300W.