12-29-2019 01:34 AM
I am a beginner in hardware acceleration of deep learning models(DNNs). I was wondering is there any open source IP accelerators from XILINX for DNNS so I can implement it on zedboard platform.
12-29-2019 01:50 AM
the short answer is NO.
The medium length answer is, all IPs are accelerators (in the sense they parallelize and offload tasks from software) so "IP accelerator" makes little sense. Now, DNNs, nice field isn't it? What is that? Algebra and ReLU functions? You can:
- Develop IPs with HLS that has libraries for these
- Use SDSoC where hardware acceleration is inferred from software
- Use PYNQ (Python on Zynq), that seems to be the trendiest way
01-19-2020 04:16 AM
@afef-86 , have you reviewed the FINN from Xilinx? While it is mainly targeted for PYNQ OS based boards [ zedboard is not in the list, check pynq.io]. But FINN could be customized for the ZedBoard and it can accelerate the ML algorithms. For more about FINN: https://xilinx.github.io/finn/about, https://github.com/Xilinx/finn, FINN Accelerated HLS library (https://github.com/Xilinx/finn-hlslib), and good paper on FINN: Paper1, Paper2. We are also working on implementing FINN based accelerated libraries in non-PYNQ supported boards.
02-08-2020 07:54 PM
02-15-2020 06:49 AM
@yoichi Actually I tried to follow this tutorial https://www.xilinx.com/video/soc/using-dnndk-for-your-custom-application-with-xilinx-soc.html
and retry the work for ZC702 board. However, I am stuck as I couldn't find the suitable DNNDK version for this board.