CNN-Processor-Module

INTRO

This project is addressed to design an Eyeris-based AI processor module. Operation uses Convolutional Neural Network (CNN) concept with the illustration as following ¹:

Operation is done between 2x2 filter matrix and 2x2 input matrix as shown in the figure below ¹:

To make the operation simpler, we first convert the 3x3 fmap input matrix to toeplitz matrix. By doing that, we translate the original convolutional operation and windowing shift into ordinary matrix multiplications as depicted below ¹:

Checklists

• Designed a simple Convolutional Neural Network (CNN) block for Eyeris-based AI processor module with pipelining and folding architecture
• Translated block diagrams into RTL codes.
• Behaviorally verified the design in Quartus and Modelsim.
• Created in 2021.

References

Footnotes

1. http://eyeriss.mit.edu/tutorial.html ↩ ↩² ↩³