Having hard time choosing FPGAs used in DNN

[Altera_Forum]

I was new to FPGAs but I'll use OpenCL for programming. I was planing to run deep learning algorithm on FPGAs.  

 Even though I will train the models on my computer, speed still need to be considered.   The algorithm that I will use is similar to super-resolution algorithm.  What kind of FPGA should I choose?

[Altera_Forum]

The best one will be the one that suits your requirements.  

What interfaces are you using? what are the Data rate requirements? what is the estimated resource usage? what external interfaces do you need?

[Altera_Forum]

Data rate should be under 10 GB/s, no special interface required.

[Altera_Forum]

Any DSP requirements? I've heard DSP is generally used for deep learning applications.

[Altera_Forum]

--- Quote Start ---  Data rate should be under 10 GB/s, no special interface required.  --- Quote End ---    With 10GB/s, you will need a special interface. You're looking at 10GB ethernet or PCIe v3.0 x16 (I think all the altera dev cards are x8).  Why such a huge data rate requirement?

[Altera_Forum]

--- Quote Start ---  With 10GB/s, you will need a special interface. You're looking at 10GB ethernet or PCIe v3.0 x16 (I think all the altera dev cards are x8).  Why such a huge data rate requirement?  --- Quote End ---    Oh, I'm sorry, that is the data rate when it's possessing the pictures, input output are pretty small.

[Altera_Forum]

I'm not sure acturally, I'll definitely using DSP functions in opencl, so maybe I should get a card with DSP chips?