YodaNN

Design of image memory in the YodaNN architecture
Block design with ARMv8 and the implementation of the image memory module
Waveform simulation of the design

Inspiration

The embedded systems evolved, and their purposes multiplied in the last couple of years. We expect our smartphones to run ML algorithms and the ASIC/FPGA entered fast the cloud processing industry. Image processing plays a role in automation and IoT. This feature can be used in medicine, security faster communication.

The inspiration was the article YodaNN: An Architecture for Ultralow Power Binary-Weight CNN Acceleration; researchers present a new architecture for embedded systems to support the performance needs in AI in the next couple of years.

What it does

Our project consists of a simulation of an FPGA board (Nexys A7) attached to an ARMv8 processor, similar to ones in smartphones or SBCs.The processor sends the pixels to the pre-decoder. This module determines in which SCM the data should be read/written. The operations execute in parallel with the main memory for optimization.

Here is a demo for running in Vivado2018.3.

How I built it

As a platform, we used Vivado Webpack because it has many testing tools and a visual interface to create RTL designs. We developed the modules in Verilog. We followed the instructions and the diagrams from the article. link

Challenges I ran into

First of all, the debugging was quite challenging because of the size of the diagram. Also, the instructions were a little unclear, but we managed to create the non-shifting version.

Accomplishments that I'm proud of

We managed to simulate a simplified version of the image memory module from the new architecture. Also, we demonstrated that the system is faster than the standard SDRAM.

What I learned

This weekend we learned a lot of things about embedded systems, memory, computer architecture, etc. Also, we improved our Verilog and circuit design skills.

What's next for YodaNN

In the following period, we will improve the research on the subject and obtain an IP authorization. Also, we could collaborate with academic research teams / R&D centers from companies to build the module.