So, I wanted to see if I could come up with an estimate of how powerful the full Skylake Iris Pro is going to be. The existing Broadwell Iris Pro has 48 execution units. Skylake Iris Pro is going to have 72 EUs, a 50 percent increase.
Though it has been released for several months, it's not easy to find benchmarks where Broadwell Iris Pro is compared against decent entry-level GPUs. Anandtech compared it against the R7 250, which is like shooting fish in a barrel. No one buys that piece of crap. What we want to know is how it stacks up to low-end GPUs that people actually care about, like the GTX 750 Ti.
Fortunately, Tom's Hardware just ran
a review of Star Wars: Battlefront where they tested it on just about every GPU, integrated or discrete, that's on the market today. This is just one game, but it is a AAA title that seems to be relatively vendor-agnostic, performing well on both AMD and Nvidia cards. At 1080p with Ultra settings and temporal AA, the Broadwell Iris Pro i7-5775C averaged 23 FPS. In comparison, the Gigabyte GTX 750 Ti OC averaged 35 FPS - a 52 percent improvement. On the red team, the Asus R9 270X DirectCU II averaged 46 FPS, exactly double Broadwell Iris Pro's showing.
Let's simplify a bit and assume that scaling for Iris Pro works exactly with the number of shaders. We can expect some architectural improvements in Skylake's EUs, but to offset that we may also be seeing eDRAM bottlenecks and/or imperfect scaling. So we'll assume these two cancel each other out, and Skylake Iris Pro has 50% better performance than Broadwell Iris Pro. This would give Skylake Iris Pro a 34.5 FPS score on Star Wars: Battlefield, almost exactly on par with GTX 750 Ti.
That would be pretty decent. The GTX 750 Ti is enough for most casual gamers, capable of doing 1080p on most games if you are OK with low/medium settings. But it won't be nearly good enough for people who want Ultra settings and/or higher resolutions, or who play the most demanding titles.
And remember, this is comparing Intel's newest (not yet released) chip against old 28nm GPUs from AMD and Nvidia. By the time Skylake Iris Pro hits, there's a good chance that one or both companies will have FinFET GPUs on the market, which should more than double the power of the previous low end (Cape Verde and GM107) at the same TDP.
If Intel's 10nm process doubles transistor density again, then conceivably Cannonlake Iris Pro could have 144 EUs, which would be about as powerful as a GTX 970. But at that point, it becomes seriously questionable whether 128MB of eDRAM is going to be enough to feed it. After all, the actual GTX 970 needs 3.5GB of dedicated RAM at 224 GB/sec. A cache can be more efficient, but that seems to be pushing it. By that time, Intel might have to resort to HBM, HMC, or some similar form of ultrafast RAM to power their best iGPUs. Anyway, Cannonlake won't be here until at least 2H 2017, and probably realistically 1H 2018 for the first desktop SKUs and 2H 2018 for Iris Pro. By that time the foundry FinFET processes will be fully mature and Intel's iGPUs will have to compete against relatively cheap ~350mm^2 chips (double the strength of a GTX 970 or more) and ~500-600mm^2 salvage chips at $650 or so (twice as powerful as a GTX 980 Ti).
TL;DR: Intel might finally come up with something playable at 1080p in the iGPU market, but AMD and Nvidia don't have anything to worry about with their discrete card businesses.