Intel has recently launched the so called CoProcessors, which are essentially oriented to parallel computing... Do you have any experience with them? Do you think they are going to compete with GPUs?
From the first one the conclusions are pssimistic for the Phi. Actually it has a problem competing with Sandy Bridge, let alone Kepler.
On the other hand the second refernce seems to say that Phi is a bit faster than Kepler on spmv, but dont trust it: the Intel code used fare more aggressive optimization techniques than the GPU code; my guess is that both architectures should yield similar acelerations for similar levels of allowed optimizations .
So, the answer to your questions will be known in, say, 10 years from now.
Phi has an advantage of easier programmabilty. But if you wish to squeeze
every flop/s from it, then the programmistic effort for Phi should be comparable to that of Kepler. There's simply no free lunch.
The fundamental question is what next? How each architecture cen develop in the future? In my opinion GPU architecture is more energy efficient by design and once it has been integrated with an ARM processor, it will easily overtake the Phi. But for this we will have to wait.
Another problem is availability of code. It is far easier to find code that will run on Phi, the question is: willl it run faster on Phi than on Sandy Bridge out of the box?
In conclusion, both architectures are experimental. None seems to be "uncondittionally better".
Thanks for answering Zbigniew! This was a very useful answer for me... I'm planning to parallelize my Matlab FEM codes in C++ and the Phi appealed very atractive to me, but your point is good... how easy it would be to squeeze it compared to GPUs.