1. Overview

In the 30th contest, participants will carry out the design through to hardware implementation, using hardware such as ZERO7 from Gigafirm.

You are free to choose the hardware. You may use ZERO7, a board made by Gigafirm, but it is not required. You can also participate using a hardware board you already have.

About ZERO7

ZERO7 is a Raspberry Pi HAT board that is stacked on the 40-pin GPIO header of a Raspberry Pi.

As shown in the architecture diagram in Figure 1, the GPIO pins of the Raspberry Pi are connected to the FPGA.

By combining a Raspberry Pi with ZERO7, you can use the interfaces of the Raspberry Pi (camera input and monitor output), and by developing LSI circuits on ZERO7, you can easily build a hardware accelerator.

Task

Details of the task will be updated at a later date. For the image size and other details, please see Section 4.

Validation images for the contest will be provided on the day. Your system reads the images from a camera, an SD card, etc. and recognizes them. In your presentation, show how fast the recognition results are output.

2. Purpose

Hardware implementation of a CNN

Conditions:

• Specified image validation data
• Real-time camera input

3. Design environment

Synopsys® Synplify Pro® /Premier

Synopsys® Design Compiler®

Mathworks® MATLAB® /Simulink®

Xilinx Vivado® Design Suite

Other logic synthesis tools

RTL hand coding (VHDL or Verilog-HDL)

Other design environments may also be used depending on your setup.

4. Overview of the CNN

Input image size

Design your CNN to accept 32×32 input images.

* The dataset and the recognition targets will be announced later.

CNN structure

For the CNN structure, please refer to the design overview of the 20th contest. The structure of the hidden layers and other details is up to you.

5. Challenge

1. Level 1: Beginner

Create a simple CNN (e.g., with a 3×3 input image) and implement it on any hardware.

2. Level 2: Intermediate

Create a CNN for 32×32 input images, and perform training and validation.

• If you use ZERO7

  Implement the CNN on ZERO7.

• If you do not use ZERO7

  Implement the CNN on any hardware board you have.

* For ZERO7 board lending and availability, please contact support@LSI-contest.com. (Mail title: Zero7 board Univ. name Your name)

3. Level 3: Advanced

Unlimited...

Any board and any architecture — anything is OK.