A massively updated architecture does not necessarily equal a massively improved architecture.
A massive update means a lot of opportunities to make mistakes.
Another thing, comparing number of EU's between different architectures is useless until you have details about both architectures. And I haven't seen such details about Cherry Trail.
There are also more opportunities for improvement. If Intel wants to challenge AMD on the GPU side for integration they have to execute on significant increases in performance, and the drop to 14nm allows extra area for GPU allocation (hence the extra EU's for both CT and Broadwell-which certainly helps), as well as CPU allocation. But even if there was no drop to 14nm, I still think architecturally there would be a decent improvement on perf/W. Broadwell-U I think will be a very solid chip performance wise, the focus is less likely to be on power consumption because the TDP is the same. Longer-term the U-series is going to pack some serious punch for high-end convertibles such as the Surface Pro 4 and others.
Another thing that interests me is the next revision of ARM v8A: Maya/Artemis, though I don't expect that revision to be in products til ~2017 (based on the 3 year duration between v8/Cortex A57 from announcement to mainstream implementation (not counting Apple, Apple may choose late 2016).
well I can already bet whatever you want that it won't beat TK1. Do you ?
I do think CT will perform higher than TK1, like I said before within 85% of HD4400 in some benchmarks. I could see it easily matching/beating the HD4200 which is still considerably faster than TK1. There will be an absolute advantage for Intel.
But then there is still the issue of Intel's 14nm Tri-Gate vs. 28nm planar. If we take process into account, for example TK1+ 20nm pln, the performance gap from where it currently is to cherry-pick 3DMark Graphics would equal the HD4200 in the i3 version of the Surface Pro 3. And on the CPU-side using 3DMark Physics (its not a bonafide CPU mark, but it is more intensive) the TK1 is already higher; higher MT geekbench than even the i5-4210y too. I can see Denver really challenging on the CPU-side. If 16nmFF+ were used for GPU, obviously I wouldn't be arguing against you. That's why on a medium to longer-term basis I think Nvidia has a real chance to make a splash with Logan+64bit-Kepler, Erista-Maxwell, and then Parker-Pascal (3x increase from Kepler-old map had Volta-4x so that's a downgrade unfortunately), but once Nvidia can close the gap node-wise to 1st gen FinFET- 16nm they'll have a much greater advantage CPU and GPU wise compared to where it is now, even if it's against Intel 10nm (depending on fin material choice). But then there will still be all of the nay-sayers trying to dismiss their capability.