03-20-2017 10:16 AM
Hello all, I wanna order an FPGA to implement the objects recognition by using DEEP LEARNING technique (e.g., Faster-RCNN). My plan is to use FPGA instead of GPU to accelerate the processing by using OpenCL programming. I searched on Xilinx website that there are several FPGA models for deep learning, such as XCVU7P, XCUV9P, etc. However, I am not quite sure what kind of FPGA are suitable for this project. Must I use a High-level Virtex FPGA? Or mid-level Kintex can also satisfy in this project? Do you have any experiences or recommendations what kind of development kits can satisfy my proposal? I wish you could give me some suggested models which are suitable for this topic. Thank you so much.
03-20-2017 10:21 AM
03-20-2017 11:08 AM
If your school is a member of the Xilinx University Program, you should contact the managing professor to ask what is already being done with Xilinx and your school. If you do not know who to ask, go to:
And send a message to firstname.lastname@example.org
There is a great deal of activity in CNN using our devices. Just about any device will do what is required. It is a matter of degree (how much, how fast) that differentiates the choices.
03-21-2017 03:31 AM
I would strongly suggest doing the design first, then buying the FPGA. Lead-time on an FPGA devkit is likely to be small - a week or less. As a result, there's really no need to order that until you're within a week of having a project ready to put on the board. By putting it off, you get lots of advantages:
- If the project is never completed (as is pretty common for ambitious university projects) then you haven't wasted your money, and there's not a board sitting in the lab for the next ten years until it gets given away as "obsolete junk".
- You can accurately select a suitable board. If your design ends up being huge, you can buy a Virtex Ultrascale+ board. If it ends up being tiny, something like the Digilent Arty will be fine.
- Better boards are coming out all the time. If it takes six months to get a decent project up and running, you might then be able to buy a ZCU102 board with production silicon, rather than the current -ES2 version.
Since you'll need SDSoC anyway, I'd probably buy something like a Zybo just for doing prototyping work and (if necessary) learning to use SDSoC. With that you can get a device-locked SDSoC voucher for $10 - making this about a fifth of the price of buying a normal SDSoC license (and you get a free Zybo!) Use that to get SDSoC up and running, build your project and get an idea of what sort of resources it's going to need, and then (once that's all done) look for a suitable devkit for the "final" project.