Historically the processor core and the Graphics have been different bitches to deal with and this is one reason, we did not see their integration until before the middle of the decade starting from 2000. Intel and AMD were processor horsepowers while ATI and Nvidia dealt with the graphics.
The things started getting interesting when ATI was acquired by AMD under the leadership of Hector Ruiz, which at that time was, criticized for being an expensive deal. It later on turned out to be savior for AMD with Brazos and Llano APUs, when Bulldozer failed to deliver.
Intel, on the other hand stared working on the integration process almost independently with its Nehalem processor. A lot of thing were new and changes of things getting out of control goes bigger as you introduce too many new things at a time. In that sense, keeping the Processor core and the graphics on separate dies made sense – if there was an error , individual dies could be debugged and fault could be isolated and fixed.
With the Westmere 32 nm processor core shrink, Intel gained confidence in the processor front and was just a step far from the inbtegrating the 45 nm graphics to its unified 32 nm. Sandy Bridge was that step which Intel launched in beginning of 2011.
The integration process is no less challenging. As you integrated the graphics and the core on the same die, validating the interaction between the processor core and the graphics becomes challenging. The debugging becomes complex – the test point between the two does not simply exist. In Nehalem as well as the Westmere die shrink you could have pads exposed out of the silicon, that could be accessed and debugged. This option was not available in Sandy Bridge.
With Sandy Bridge the cost and the power improved, but for engineers it was harder to debug since there were fewer places to test.