GPU chiplets are in pipeline. GPU itself is embedded in infinity fabric already since Vega. For certain tasks Chiplets are no problem. For other tasks, access speed to memory attached to other GPU chiplet(s) may prove to be an issue. Using I/O hub as with Zen2 may prove to be problematic as GPUs use magnitude higher bandwidth than CPUs. Complexity may prove to be too much. We'll see with 8-channel 8chiplet Zen2. But that uses rather big I/O die and that has to be paid too. It may start to be useful in case AMD goes for 5 times 10 Dual-CU (20 CUs per chiplet = 1280SP in 1st Navi configuration) => 6400SP. But monolithic GPU with 5120SP may still have reasonable size (price) on 7nm to offset cost of interposer for chiplets and huge I/O die.