Description
Deep Learning in Hardware
Background:
So far, we have talked about various topics in the lectures, including finite precision analysis and DNN
training, such as LMS (least mean square) and SGD (stochastic gradient descent). In Homework 1, you
used Wiener-Hopf method to compute the optimal weights and MSE for a 3-tap linear predictor. In
Homework 2, you used SGD to train your weights and observed that they approached the optimal weights
and so was the MSE. You also pipelined your architecture such that it achieves lower critical path delay.
In Homework 3, you experimented with quantized gradient update, where you limited the precision used
for backpropagation. You should have built background knowledge on how to build and train a network
with finite precision and how to construct hardware mapping such that it has better performance.
Project Description:
In this mini project, your goal is to implement the 3-tap linear predictor from Homework 2 in hardware.
Your task is to design the 3-tap linear predictor in fixed-point such that its MSE is within 0.5dB of the
optimal MSE. You will be using Python as your software prototyping language and Verilog as your
hardware prototyping language.
You will then have to implement this 3-tap fixed-point predictor in Verilog or SystemVerilog using
Vivado, or any other applicable software that you want. All your implementation should be in fixed point
and you need to be careful about the precisions in your design, namely the bit growth during computation
and setting your output back to the desired precision. Once you finish coding and verify your functionality
through simulation, i.e., making sure the Verilog output matches the output of the fixed-point Python
code, you need to synthesize it and report the resource utilization, design latency, design energy, and
schematic. Interpret the numbers available to you in the context of your design.
Deliverables:
● Type report for you hardware and software summary, including
○ Optimal solution and software design strategy, i.e, quantization.
○ Hardware design strategy and report of the resource utilization, timing, design power, and
schematic.
● Python Code for your software prototyping
● HDL Code for your 3-tap fixed-point predictor (Verilog/SystemVerilog)
● Behavioral simulation that demonstrates your design correctness. You should show sufficient
time steps and values to show how your design works.
Timeline:
● Release: 10/23/
● Due: 11/06/
Grading:
● 10% Software Prototyping
● 60% Hardware Prototyping
● 30% Clarity and exposition of your written report.
