I don't believe the Iris Pro is as fast as the 650M/750M, close, but not quite.
It is, actually. Have a look at notebookcheck. In most cases it is either on par or marginally slower.
Still a solid effort from Intel, but to get any faster they are going to need to include more CU's, pipelines etc. No amount of fancy architecture will change that, then you run into the problem of having a massive and rather expensive die.
True, but again: it is a very capable GPU, and the architecture isn't where the major bottleneck is.
DDR3 is a problem as GPU's need large amounts of memory bandwidth to feed them - just look at DDR3 vs GDDR5 versions of the same chips. The XBOne is going to run into severe issues with this over the coming years, using 32MB of ERAM aka L4 Cache is too small and inadequate for gaming, hell Iris Pro even has 128MB!
That is exactly what I said: the performance of iGPUs is EXTREMELY hindered by the RAM. As soon as we get faster RAM, their performance will skyrocket.
Intel's only solution is to use GDDR5 on die and that is prohibitively expensive, or OEM's like Apple and Dell include it themselves.
They can't use GDDR5 on die due to its poor latency. It's inefficient as a cache. The best solution on the horizon is the HMC, that has both low latency and high bandwidth, but that is a very young technology and will probably not appear in mainstream computers as soon as we'd like it to.
I still think the best solution is what we have currently, with a strong iGPU along with a solid dGPU. Looking forward to seeing how Maxwell performs, as my current 750M runs everything pretty well. If it received a 50% boost I would be extremely happy.
I can't quite agree with you. This combo has its flaws, and as integrated solutions will likely keep closing the performance gap, the extra heat and power draw, generated by dGPUs will be getting less and less worthy of the hassle. However, at the moment it has no alternatives if you actually need all that graphics power.