A big reason for the the monolithic die is that without it's own pool of memory, the iGPU is entirely dependent on the CPU's memory controller and having the IGP on a different die, even on the same package, is a significant bottleneck.
Back in the Clarksfield days, Intel had the IGP and the memory controller on a separate die as something of a compromise.
HBM2 is probably the first time they've had a memory solution with enough speed and density to make a decent IGP practical with a multi-die setup that also didn't compromise the CPU performance.
Broadwell C CPUs could have gotten better performance if probably they used current KBL iGPU?
later intel could use HMC(DDR5) based memory if bandwidth is a requirement
Edited by PontiacGTX - 4/4/17 at 7:16am