I hear knights landing (whichever the latest is) will have 72 cores and a bootable OS.
haven't looked into it enuf, short a few grand right now.

google is your friend,
I don't remember the articles I've been reading about it.
I don't think Knights corner will work,
but if I'm remembering correctly knights landing
(the latest one) should do the trick.
I want one too,
but price supposed to be around $4000 USD.

Edit: http://www.servethehome.com/intel-knights-landing-details-emerge/

Klik,
I have to disagree about the power consumption.
If it runs like a CPU (which is a big IF)
then my latest I7 5820 6 core build with radeon 7950,
uses more than 200 watts.
If this other thing is doing terraflops then it kicks the 5820
to the curb. (less than a 100 Gflops)

Trying to compare Xeon Phi to existing CPU or GPU doesn't really work as it is something of a combination. On one hand each Phi core it is much "smarter" (has a simplified x86 instruction set available) than a GPU core but on the other hand those same cores were never intended to fully have the same functionality as a modern CPU core. The balance of which is why they have (only) 61 cores. SP and DP specs are not indicative of anything unless you can say what the software you are running depends on - do you know?

Probably one of the largest factors in any discussion of efficiency between CPU / GPU / Phi is if / how well the problem itself can be solved in a parallel manner. Think about it this way - there are very few BOINC GPU projects but lots and lots of CPU ones - that is a combination of the skill of the scientists' programing abilities (no disrespect intended - they often hold doctorates in their specific field but only program because they have to) and the fact hat most of these "problems" are not solvable in a parallel manner to begin with. This is not easy stuff.

There is another case for the Phi - and that is in theory it should be "easier" for developers to port code between CPU and Phi (remember there is at least some level of commonality on the x86 knowledge) . A great example of this is what *could* be done with the Rosetta application (while the outcome is speculative, the background information is from the real world). Rosetta does not use any advanced cpu vector instructions (not even SSE) so it would likely be a relatively easy port - we're not talking about changing the programming model to be parallel, likely just a wrapper so that a cpu could manage kicking off multiple instances on the Phi. While this is the least beneficial of optimizations, you could be running 61 WUs at a time. To bring it closer to home, up until a few years ago we were running a genome project here on WCG that used that app!!! Trying to make that app run on a GPU would not work at all, you would have to write an entirely new program based on entirely different instruction set - as a programmer I can say that is a huge undertaking and not something to be taken on lightly.

apples + oranges + plums = enjoy the full cornucopia