SEU-MSLab / PRVTDNN

The Chisel design of Polyphase Real-Value Time-Delay Neural Network (PRVTDNN).

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Hardware Implementation of Polyphase Real-Valued Time-Delay Neural Networks

Introduction

This repo contain the source code of Polyphase Real-Valued Time-Delay Neural Networks Implementation designed in Chisel.

System Architecture

The architecture include

  • Systolic array: to do the matrix multiplication.
  • Bias and activation module: to add biases and apply activation functions, we implement three, ReLU, Sigmoid and Linear.
  • Parameter loading module: to propagate weights to systolic arrays and biases to bias and activation modules.
  • Pipeline dataflow: the DPD is a real-time problem, so multiple systolic arrays are cascade to make a fully pipeline neural network implementation at the expense of more area.

The system architecture

Performance

The following experiment result using 40 MHz 5G-NR signal with a 3.5 GHz GaN power amplifier. The test board is AMD VCK190 evaluation board.

# Network Configuration Activation function Number of parameters Type NMSE (dB) ACPR (dBc) DSP URAM LUTs Registers
N/A w/o DPD N/A N/A N/A -25.35 -35.98/-33.91 N/A N/A N/A N/A
1 [12,10,10,8] ReLU 328 S/W -31.35 -43.47/-43.78 N/A N/A N/A N/A
H/W -31.28 -43.29/-43.28 300 0 9034 11082
2 [20,10,10,8] ReLU 408 S/W -32.14 -44.07/-44.77 N/A N/A N/A N/A
H/W -31.25 -43.91/-44.06 380 0 11327 13549
3 [20,10,10,10,8] ReLU 518 S/W -32.51 -43.49/-44.45 N/A N/A N/A N/A
H/W -32.34 -43.36/-44.33 480 0 14281 16877
4 [12,10,10,8] sigmoid 328 S/W -30.10 -43.95/-44.15 N/A N/A N/A N/A
H/W -29.52 -43.97/-44.18 300 19 9035 11334
5 [20,10,10,8] sigmoid 408 S/W -34.57 -46.07/-46.58 N/A N/A N/A N/A
H/W -34.03 -45.19/-46.07 380 19 11328 13787
6 [20,10,10,10,8] sigmoid 518 S/W -34.86 -48.06/-48.38 N/A N/A N/A N/A
H/W -35.28 -47.63/-47.89 480 29 14313 16830

Prerequisites

  • JDK 8 or newer. You can install the JDK as recommended by your operating system, or use the prebuilt binaries from AdoptOpenJDK.
  • SBT. SBT is the most common built tool in the Scala community. You can download it here.
  • Firrtl. Firrtl is the tool to convert firrtl to verilog file. You can download it here, and put it to your environment variable $PATH.

How to get started

git clone git@github.com:SEU-MSlab/PRVTDNN.git
cd PRVTDNN

Generate the Verilog file with

sbt run

Run the test with

sbt test

You should see a whole bunch of output that ends with something like the following lines

[info] Tests: succeeded 1, failed 0, canceled 0, ignored 0, pending 0
[info] All tests passed.
[success] Total time: 5 s, completed Dec 16, 2020 12:18:44 PM

About

The Chisel design of Polyphase Real-Value Time-Delay Neural Network (PRVTDNN).

License:MIT License


Languages

Language:Scala 100.0%