You say that ARM chips are 10x slower than Haswells, but the difference isn't actually that large anymore and for Apple's A7 it's around 3x versus a normal desktop chip like the 4770 and only 1.7X versus the 4250U you'd find in a Macbook Air[1]. Diminishing returns is a constant factor in engineering and I'd expect it to be very hard for Apple to make up the remaining distance between themselves and Intel, but we shouldn't exaggerate how large that difference is. And it is possible that some hypothetical A8 could be made competitive with Intel in the 15W range due to not having to make the compromises required to hit 4 GHz.
I do agree that the ~10W you'd save by using an A7 versus an 4250U in a Mackbook Air would be a false economy.
EDIT: To explain that bit about targeted power ranges a bit more, to hit high frequencies at a given process node (like 45 nm or whatever) you need to break up your logic into a fairly high number of simple stages separated by clock driven latches. A core that doesn't worry about hitting high frequencies can divide up it's logic into fewer stages, simplifying design and requiring fewer transistors for latching.
I think GeekBench is a poor benchmark, since it's measuring peak CPU performance in a narrow domain, but the table you cited shows that the 4770 is 6.5x faster than the A7 (16774 vs 2564). Even at the same frequency and core count, Haswell is 1.8x faster in the benchmark.
Whatever the exact number, it doesn't invalidate the point I was making: Switching to ARM will ruin performance for x86 programs, which need to be emulated. ARM has to be significantly faster than x86, or else the emulation overhead will make users hate the transition.
This is a huge challenge for the Mac Pro and rMBP, which are all about performance. Generally, a given microarchitecture can only scale TDP by a factor of 10 by changing frequency and number of cores. Satisfying the pro line would require a new microarchitecture. Considering an x86 to ARM transition would take 1-2 years, that's asking a lot of Apple's chip designers. If Apple was going to go through all the trouble of building a new ARM core for high-performance, high-TDP applications, they might as well build their own x86 CPU. It would save a transition and the necessary emulation overhead. They could roll it out for the MacBook Airs first, then address the Pro line at their leisure (or not and just use Intel CPUs).
> for Apple's A7 it's around 3x versus a normal desktop chip like the 4770 and only 1.7X versus the 4250U you'd find in a Macbook Air
That would still be noticeably slower for native software, and unbearable for anything run under emulation. It would also be very painful for use cases such as people wanting occasional Windows applications or Linux VMs.
Right, using an A7 in a Macbook would be totally crazy. The idea is that Apple might create an A8 that's about as fast as the current Macbook Air and use that. Getting as much performance increase going from A7 -> A8 as they did going A6 -> A7 strikes me as implausible, and it would probably be a bad idea to switch architectures anyways. But it's not totally crazy.
I do agree that the ~10W you'd save by using an A7 versus an 4250U in a Mackbook Air would be a false economy.
EDIT: To explain that bit about targeted power ranges a bit more, to hit high frequencies at a given process node (like 45 nm or whatever) you need to break up your logic into a fairly high number of simple stages separated by clock driven latches. A core that doesn't worry about hitting high frequencies can divide up it's logic into fewer stages, simplifying design and requiring fewer transistors for latching.
[1]http://www.computingcompendium.com/p/arm-vs-intel-benchmarks...