Use EdgeBoard Lite as a general Zynq development board
EdgeBoard Lite (a.k.a. FZ3) is an FPGA-accelerated embedded AI computing board, launched by Baidu Brain in 2019. As part of the Baidu AI solution, it can quickly depoly the neural network models.
The board is based on Xilinx Zynq Ultrascale+ ZCU3EG MPSoC, which can achieve 1~2 TOP/s with an average power of 5~10 W. It also has the following peripherals:
- 2GB/4GB DDR4 RAM
- 8GB eMMC Flash
- 256Mb QSPI Flash
- PCIe x1 connector
- MIPI camera connector
- BT1120 HDTV connector
- 1000Mbps Ethernet
- TF-card slot
- USB Hub (1 x USB 3.0, 1 x USB 2.0)
- Mini DisplayPort (miniDP)
- USB UART
In addition to the rich peripherals, what matters more is that it only costs ~1000 CNY (about 140 USD)! It is almost the cheapest Xilinx Zynq Ultrascale+ development board. Another Zynq ZCU3EG board Ultra96, much loved by developers, is priced at 249 USD. In the Chinese market, Ultra96 even costs more than 2200 CNY, more than double of EdgeBoard Lite.
Unfortunately, Baidu sells the EdgeBoard as its AI solution instead of a general-purpose development board. Therefore, it is a bit difficult for us to obtain full official support (You can send emails to Baidu to request limited documents).
This repository contains materials that help us use EdgeBoard as a general Ultrascale+ Zynq development board.
Here are some documents you may need, and you can find them in the ./docs folder. It should be noted that most of them are in Chinese, and currently do not have an English translation.
- FZ3A Hardware handbook, provided by ALINX. Its online version can be also found on Baidu.
- FZ3B Hardware handbook, provided by ALINX. Its online version can be also found on Baidu.
- FZ3A PCB-level circuit schematic, provided by ALINX
- FZ3A IO pin assignment, provided by ALINX
Most of above documents are from ALINX, the manufacturer of this board.
Documents about how to run EasyDL/PaddlePaddle models on the EdgeBoard Lite will not be listed here. You can visit its product website to access them.
I have ported the PYNQ framework to EdgeBoard Lite, and provide the prebuilt image file. All the necessary source code are also open sourced.
They are in ./board_files. Please read its inside README for more details.
The prebuilt image files are compiled based on PYNQ v2.7. The images are put on Aliyun Cloud Drive:
- FZ3A: Link,
- FZ3B: Link (thanks to @changhai0109)
The folder ./pynq is board specification files, which are necessary for PYNQ compilation. If you want to build from scratch, you can read this post on my blog (sorry that it is in Chinese).
Except for those already noted in the text, other documents written by myself are released under the CC-BY 4.0 and the code is released under the MIT license.
This repository is created and maintained by me personally and has nothing to do with Baidu.
- @zhutmost created this repository.
- @changhai0109 contributes the FZ3B compilation flow.
Thank them for their contributions.