Have you guys at PS ever looked at supporting this guy? http://www.intel.com/content/www/us/en/high-performance-computing/server-reliability.html
This one has 61 cores and is passively cooled: http://www.amazon.com/Intel-Xeon-Phi-7120P-Coprocessor/dp/B00FKG9R2Q/ref=sr_1_1?ie=UTF8&qid=1406043069&sr=8-1&keywords=intel+phi
Just by dumb luck I know the brother of the general manager in this division. Also, just by dumb luck, the parallel programming language I've been developing is called ϕ (as opposed to its English phonetic spelling "phi").
The advantages of this strategy over a GPU is that the cores are x86 and not something different. They (apparently) have the same direct access to system main memory as do the host CPU cores. Of course AMD's API for their GPUs "Mantle" also gives this direct access to system resources but the cores are not x86. The down sides of using the Intel product are that (firstly) you can't just recompile code written for the host processor cores. I think you have to use Intel's compiler and there is some call overhead that is specific to their system. Secondly, significant cost will push it out of budget range for most of your users.
It's an interesting technology. But as you say it is currently cost-prohibitive for most users, and still very much a niche product. If it gains more adoption it's something we could consider more strongly.
- Oshyan