What happened there? They could have just switched wafer production 1:1 from AD103 TO GB203 but that clearly didn't happen.

Problem is that the 4080s/4090s are discontinued already. Which leaves nothing in that performance range. I don't understand this strategy other than that they really don't care about the consumer market anymore. The 5080 doesn't even have a bigger die and it's same process node.

Note on the subgroup thread to thread id mapping thing: It's not even guaranteed that all threads in a subgroup are active at the start of the shader. Intel actually has this behavior for whatever crazy reason.