I've asked this question (what about OpenCL support for Xeon Phi x100 and x200?) several times at Intel meetings, and never got a satisfactory answer... it seems Intel is promoting support for OpenCL on every product (desktop CPUs incl their GPUs, server CPUs, FPGA chips) except for the Xeon Phi's.
In short:
- The Xeon Phi x100 range supported (supports) OpenCL but performance is mediocre; rewriting your code using OpenMP is more effective. Also, note that OpenCL support for the x100 range has officially been ended.
- The Xeon Phi x200 CPU series (non PCI-e) behave like regular CPUs yet they are not supported using Intel's own OpenCL CPU driver, which does support Atom, Core i5/6/7, etc.
- With a lot of trickery, it is possible to get OpenCL code running on the Xeon Phi x200 CPU, using either AMD's (!!) APP kit, which is very slow, or using an older version of the Intel OpenCL driver, which does not support AVX2 or AVX512 (i.e. it's also slow)
- The Xeon Phi x200 PCI-e card have gone AWOL as per July 20th or so, so it's impossible to tell if OpenCL support for those cards is achievable (though it would never be supported by Intel)
This all seems to be a software design decision by Intel, not a hardware issue. I'd be very interested to talk to an Intel OpenCL Developer on some hidden tweaks to get AVX512 working in the OpenCL autovectorizing engine for the Xeon Phi.
PS a post about NVidia supporting also seems to have gone AWOL; my note on that would be: NVidia's support for OpenCL is abysmal; yes they support it, but performance is very poor. Note also that they've thrown out all OpenCL SDK stuff from both the driver and the Cuda SDK.