Actually, I fear for NVIDIA's HPC market as well. Kepler 2 in my eyes is not well suited for GPGPU, they've rather optimized it for the gaming market in order to go after AMD. I.e. they introduced much more cores, but those cores now have less cache and register resources available to them. Less registers, especially, is deadly - this has already been a limiting factor for many HPC applications.

Meanwhile Intel pushes out their Xeon Phi boards with huge memory bandwidth and OpenMP support. If there's not some kind of surprise in terms of performance either on Intel's or NVIDIA's side I see dark clouds coming for them.

Well Intel recently introduced its Xeon Phi co-processor, which according to them offers about the same performance as Nvidia's chips, while not needing to code in CUDA. But I never trust Intel's marketing anyway because they always seem to exaggerate in some way or be misleading on purpose, so we'll see how that goes. Plus, I'm interested to see what comes out of Nvidia's Project Denver or whatever they are calling it now (Nvidia's 64-bit custom ARMv8 SoC for HPC and servers, which should arrive in 2014).

>But I never trust Intel's marketing anyway because they always seem to exaggerate in some way or be misleading on purpose, so we'll see how that goes.

I think you can safely remove "Intel's" from that sentence and it remains as accurate, or perhaps more so. I really don't mean to hate on marketing, but their job is to sell a story. Good marketers walk that fine line between outright lying and carefully walking the "happy path" that makes whatever they are selling the cure for what ails you and fails to mention the problems. In general once you dig into things you almost always start finding the trade-offs/drawbacks that marketing "forgot" to mention. Thomas Sowell was right when he said: "There are no solutions ... only trade offs".