Undergrad Research Project - Neural Networks on FPGAs

Spring 2017

Anand Prasanna
Raj Rajkumar
Project description

Objective: The objective of this project is to investigate how Neural Network based Deep Learning Systems perform on Field Programmable Gate Arrays(FPGA). Specifically, this project will study Convolutional Neural Networks(CNN) and the performance vs power tradeoff that can be achieved by using specialized hardware versus more general platforms like GPU’s.

Method: We will first investigate and understand current approaches to implementing Convolutional Neural Network architectures like Eyeriss from MIT and EIE from Stanford. We will then create software that allows for initial prototyping and interfacing between a CPU and the FPGA. Once we have the initial architecture sketched out, we will then implement the neural network to maximize parallel operations on the FPGA.

Anticipated Results: We plan to initially have a subset of the CNN architecture(e.g. Convolutional Layers or Fully Connected Layers) implemented on an FPGA. Ideally, by the conclusion of this project, we would have a complete end-to-end deep neural net accelerator implemented on an FPGA.

Return to project list