FPGA Artix7 Basys3 Deep/ Machine learning

[Tim.O]

Are there any introductive hands-on web pages or articles on Deep/Machine learning on FPGA Artix7 (not ZYNQ or PYNQ) ?? I'm doing deep learning using Tnsorflow or KELAS, and now, I'm thinking of performing it on FPGA Artix7. But, I do not have any clues to start it. If someone knows useful information, could you tell me. 

Best regards.        

Edited January 19, 2021 by T.O

[JColvin]

Hi @T.O,

Digilent does not have any materials on machine learning, but you might want to look into these pages I found here and here.

Thanks,
JColvin

[Tim.O]

Thank you for your comments. I'll check web pages you linked!

 

Personally, I was guessing the procedure.

-Using any D.L. framework, create a weight file that matches the real configuration (FPGA channel or I/O) 

-Rewrite it to C++ and put it on FPGA with Micro-blaze (soft CPU)

-Based on FPGA input: High/Low, FPGA provides High/Low following the weight file. 

But, if #of FPGA I/O is not sufficient, a couple/few of FPGA must be used. In that case, how is communication among them done....

 

Anyway, I have to investigate it. PYNQ is probably helpful to brush my knowledge up...

[zygot]

A much easier path would be to choose an FPGA platform that has the infrastructure that you want in place. You should at least look at the Terasic OpenVINO (formerly C5P ) starter boards. No soft processor. Just PCIe, the FPGA application and a PC application. Windows and Centos 7 drivers are included. There's plenty of DDR memory o the FPGA board. I've written software applications for Centos and Win10 for these boards so I know that they work. Yes, you will have to do your FPGA development using Intel tools. If you decide to go this way the GT version is the one that you want as it supports PCIe Gen 2. The GX version only PCIe Gen 1.

Selecting a platform and trying (hoping) to make it do what you want is always a lot more work than selecting a platform that is designed to do what you want.

If there were a similar options for Xilinx users in the price range I'd have tried it and would offer it as a possible candidate for a project similar to yours. I don't believe that there is but would love to be proven wrong.

Edited January 20, 2021 by zygot

[ucc]

Hi, I have same question now , have you already solved your question ?