Adaptive deep learning hardware

Before start:

this is a project with pre-built hardware confs that you can use to in Vitis adaptive platform. For more info about how to run this project. please visit https://www.hackster.io/378085/adaptive-deep-learning-hardware-for-video-analytics-f8d064

What is this project for?

We want to build a convenient hardawre switch framework on embedded system to change the hardware functions dynamically to satisfy the requirements of functionality, performance and power consumption in different AI scenarios.

The platform comes with pre-built hardware configurations, you can switch the hardware configurations as you want.

What do you need to run this project?

hardware

• ZCU104 development board
  • The design is based on zcu104. You can follow the link to build configurations for zcu102/106 or other zynqmp platforms
• SD card (16GB or more)
• USB camera

software

• Vitis 2020.1
• Vivado 2020.1
• Petalinux 2020.1
• Vitis AI 2020.1

What do we have now? (2020/11 version 0.99)

The project is now in early access. We have provided some configurations supporting both DPU and VCU, but user need to run management program to trigger switch manually. Auto switch feature will be integrated in next version.

The current configurations are listed as follow：

conf	number of VCU	number of DPU	frequency of DPU(s)	fps	power
0dpu_acc	1	0	/	/	/
1dpu_l	1	1	150MHz	450	7.991W
1dpu_m	1	1	200MHz	488	8.875W
1dpu_h	1	1	300MHz	742	10.871W
2dpu_ll	1	2	150:150MHz	811	10.859W
2dpu_lm	1	2	150:200MHz	977	11.99W

Note: the fps means the performance running facedetect with densebox_320_320

How to run

refer to the link and follow the instructions: https://www.hackster.io/378085/adaptive-deep-learning-hardware-for-video-analytics-f8d064

ALL-IN-ONE image: https://drive.google.com/file/d/1axsFStKz79R-mJ8R3z9-eJ5-pwGJFTEg/view?usp=sharing. Download and flash it to your SD, the design will be ready to work.

Hardware switch

Running facedetect