Broadwell had a GT3 variant, with 48 EUs. However, in Skylake, there was a GT4 variant with 72 EUs.
According to what I recall, the Iris Pro w/ eDRAM were possibly even bottlenecked on accessing it. The eDRAM bandwidth was much lower than you'd expect.
Acording to tests, I think here at tomshardware, they were upward 150% faster than (2.5 times) than their none edramm variants.
And yes, they used more power, than their counter mobile parts, but it smashed all in the tests I saw it at, at integrated gpu level.
Going on package 64-128MB level cache/ram is defo gonna be faster than accessing your ddr4 memory; or, said in other words: The Increase in EU from 48 to 72, couldn't alone make up for that 'giant' improvement over the none edram versions, from the tests I saw.
My point is just, that a Iris variant beats out a HD variant, ain't that impressive at all, since they' always done that, with huge margins.
NB: The listed memory speeds at wiki, is the speed of the ddr4 memory, not the edram
NBB: Im' quite sure i've read Intel was opening up for using the edram for cpu, in scenarios where the igp couldn't use all of the level4 cache(which technical was a cache for the igp, at first, but promised also cpu later..). I never read that this came true however..