Another point, GPGPU computing is not the magic bullet few people thinks, a key difference on GPGPU is on Thread/memory Access on massive parallel compute environment every thread can communicate with all other threads and request some data or access their memory registry (as could be running a process distributed along 100 Pcs on a network) GPGPU restricts memory/communication to pools of threads or thread groups outside that you need to wait for this pool or group to read an result or write new data, it's like having a distributed process along 100 pc, but on groups of 10, where only those process inside each group can access the data from the other 9 pc and only at end can relay to the thread master, this is very restrictive on what you can do with an GPU to process some data.
That why there is the Xeon Phi, it's more like a bunch of computers on a network inside a single chip, notwithstanding the best Xeon Phi is slower on TFlop/s than the best GPU, it's more likely you can run an huge process on a Xeon Phi than on an GPU, since you don't have major restrictions on the algorithm than you have on a tipical multi threaded application.
I tried Cuda/OpenCL but finally I ended on Xeon Phi for my projects, so I's unlikely I'll ever use a pu again unless the algorithm enable it, but i prefer to invest on Xeon Phi or high core Xeon than GPGPU.
Of course I follow developments on the GPGPU market since the mentioned restriction it's something that could be eliminated soon or later.
That why there is the Xeon Phi, it's more like a bunch of computers on a network inside a single chip, notwithstanding the best Xeon Phi is slower on TFlop/s than the best GPU, it's more likely you can run an huge process on a Xeon Phi than on an GPU, since you don't have major restrictions on the algorithm than you have on a tipical multi threaded application.
I tried Cuda/OpenCL but finally I ended on Xeon Phi for my projects, so I's unlikely I'll ever use a pu again unless the algorithm enable it, but i prefer to invest on Xeon Phi or high core Xeon than GPGPU.
Of course I follow developments on the GPGPU market since the mentioned restriction it's something that could be eliminated soon or later.