... is it possible to use C++ AMP in the application
I have some experience with CUDAfy.NET and investigated some ideas around integrating GPU processing into FSB a while ago. My belief is that while it is probably possible, the time involved is unlikely to provide any outcomes that are better than what is possible with the changes that are currently planned for the application.
There is a big overhead in moving data in and out of the GPU in order to offload the calculations, and due to the number of calculations performed during a generator process it is likely that the processing would actually be slower.
As an example, working with 50,000 bars with only 100 FP calculations per bar in a generator run that produces 5,000 strategies would perform 25 billion FP calculations (and there would likely be a lot more than 100 FP calculations per bar). In order to process these on the GPU, a massive amount of data would needed to be moved back and forward due to the differences in how a CPU -v- a GPU operates.
Assuming that the above process takes 4 hours on your hardware, it is likely that the overhead involved in moving data to and from the CPU to the GPU and back would increase this time, given that now we only need to perform the FP calculation and not shift data around.
My understanding is that popov has a v 2.8 in the works for Q1 next year, with one of the big changes being the ability to multi-thread the generator (and maybe/hopefully the optimizer) process. In this event, the addition of more CPU cores would provide a more cost effective way of speeding up the analysis process, particularly given the amount of work that would be needed to integrate CUDYfy (or another library).
There is also the obvious issue that there would likely be a very small number of people who have access to GPU capabilities, so the amount of work involved wouldn't necessarily provide any benefit to the wide user community.
Just my thoughts on the topic.