There will be a point where APUs will make dGPUs obsolete. Their advantages are huge and it’s just the raw power and bandwidth that needs to catch up. Things like lower latency between CPU and GPU, and the big one: being able to use a shared address space. I don’t think even today’s APUs generally do that and instead set aside some system RAM to act as video RAM, but that setup involves a lot of copying data back and forth between video memory and system memory. If they both can just access the same memory space, that no longer needs to happen at all.
So it’s just a matter of fitting more compute cores on that package (which isn’t limited by chip size with chiplets) and scaling up the memory bandwidth until those advantages above reach parity with a dGPU.