We'll see if it actually gets that. On Ampere, rather than crashing the memory will generate more and more errors. Actually it probably was somewhat true in the previous generation too, but to a much lesser degree.

I have 3 Polaris cards running. 1 card is weaker in this aspect than the rest. None of the cards get memory errors nor do they get incorrect shares. But the weaker card needs the memory clock to be backed down little bit further otherwise the effective hash rate drops compared to the reported hash rate.

I've seen other systems on Etherscan where the effective hash is consistently 5-7% lower than the reported hash rate. This is what happens when the card is pushed too hard but not so much that monitoring utilities are able to see the problems.