Isn't this a sign that AMD's test suites are incomplete?
If they implement a feature correctly then break it in the new design, it suggests their test suites didn't exercise it properly, otherwise (you'd think) they would have caught the regression in the new design and fixed it before release.