paper.tex

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% 서울대학교 데이터마이닝연구실 구성원들의 박사학위논문 작성을 위해 아래 저작자의 자료를 일부 수정하였습니다.
%% Author: zeta709 (zeta709@gmail.com) 
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% 2018-01-08 산업공학과 데이터마이닝 전공을 산업공학과로 수정하고, 영어 석사 논문 작성 시 편리한 작업을 위한 추가 수정 및 팁 작성 by 문지형 jhmoon@dm.snu.ac.kr

\RequirePackage{fix-cm} 
% oneside : 단면 인쇄용
% twoside : 양면 인쇄용
% ms: 석
% phd : 박사
% openright : 챕터가 홀수쪽에서 시작
\documentclass[oneside,ms]{snuthesis_utf8}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% 목차 양식을 변경하는 코드
%% subfigure (subfig) package 사용 여부에 따라
%% tocloft의 옵션을 다르게 지정해야 한다.
%\usepackage[titles,subfigure]{tocloft} % when you use subfigure package
\usepackage[titles]{tocloft} % when you don't use subfigure package
\makeatletter % don't delete me
\renewcommand\cftchappresnum{Chapter~}
\renewcommand\cftfigpresnum{Figure~}
\renewcommand\cfttabpresnum{Table~}


\usepackage[pdftex,bookmarks=true]{hyperref}
\usepackage{tabularx}
\usepackage{array,multirow,graphicx,rotating,booktabs}
\usepackage{caption}  %subfigure
\usepackage{subcaption}  %subfigure
\usepackage[round, sort, numbers]{natbib} %reference style
\newcommand\mycite[1]{[\citenum{#1}]}
\newcommand\myauthor[1]{\citeauthor{#1} (\citeyear{#1})}
\usepackage{pbox} % line break in table
\usepackage{adjustbox}
\usepackage{footnote}
\usepackage{color,soul}
\usepackage{colortbl}
\usepackage{amsmath}
\usepackage{float} % position here
\usepackage{lscape} % for landscape page
\usepackage{kotex}
\usepackage{listings}

\makeatother % don't delete me
\newlength{\mytmplen}
\settowidth{\mytmplen}{\bfseries\cftchappresnum\cftchapaftersnum}
\addtolength{\cftchapnumwidth}{\mytmplen}
\settowidth{\mytmplen}{\bfseries\cftfigpresnum\cftfigaftersnum}
\addtolength{\cftfignumwidth}{\mytmplen}
\settowidth{\mytmplen}{\bfseries\cfttabpresnum\cfttabaftersnum}
\addtolength{\cfttabnumwidth}{\mytmplen}
%% 목차 양식을 변경하는 코드 끝
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% 다른 패키지 로드
%% http://faq.ktug.or.kr/faq/pdflatex%B0%FAlatex%B5%BF%BD%C3%BB%E7%BF%EB
%% 필요에 따라 직접 수정 필요
\ifpdf
	% \input glyphtounicode\pdfgentounicode=1 %type 1 font사용시
	%\usepackage[pdftex,unicode]{hyperref} % delete me
	%\usepackage[pdftex]{graphicx}
	%\usepackage[pdftex,svgnames]{xcolor}
\else
	%\usepackage[dvipdfmx,unicode]{hyperref} % delete me%
	%\usepackage[dvipdfmx]{graphicx}
	%\usepackage[dvipdfmx,svgnames]{xcolor}
\fi
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
%% \title : 22pt로 나오는 큰 제목
%% \title* : 16pt로 나오는 작은 제목
\title{Practical Partial Row Activation\\ for 3D Stacked DRAM with Applications to Deep Learning Workloads}
\title*{3D 적층 DRAM을 위한 실용적인 Partial Row Activation 및 딥 러닝 워크로드에의 적용}

\titlen{Practical Partial Row Activation\\ for 3D Stacked DRAM with Applications to Deep Learning Workloads}

\author{Namho~Kim}
\author*{Namho~Kim} % Same as \author.
\authorn{Namho~Kim}
\phonenumber{010-5120-5105}
\studentnumber{2017-22393}
\advisor{Jae~W.~Lee}
\advisor*{Jae W. Lee}
\advisorn{Jae~W.~Lee}
\graddate{February~2019}
\submissiondate{January~2019}
\submissiondaten{January~20th,~2019}
\approvaldate{January~2019}

\committeemembers%
{Jihong Kim}%
{Jae W. Lee}%
{Jaejin Lee}%
{교 수 님}%
{교 수 님} %

%% Length of underline
\setlength{\committeenameunderlinelength}{5cm}

\begin{document}
\pagenumbering{Roman}
\makefrontcover
\makeapproval

%agreement page
\cleardoublepage
%\makeagreement

\cleardoublepage
\pagenumbering{roman}

\keyword{DRAM Architecture, GPU, Convolutional Neural Network, HBM2, Power Efficiency}
\begin{abstract}
GPUs are widely used to run deep learning applications. 
Today's high-end GPUs adopt 3D stacked DRAM technologies like High-Bandwidth Memory (HBM) to provide massive bandwidth, which consumes lots of power. 
Thousands of concurrent threads on GPU cause frequent row buffer conflicts to waste a significant amount of DRAM energy. 
To reduce this waste we propose a \emph{practical} partial row activation scheme for 3D stacked DRAM. 
Exploiting the latency tolerance of deep learning workloads with abundant memory-level parallelism, we trade DRAM latency for energy savings. 
The proposed design demonstrates substantial savings of DRAM activation energy with minimal performance degradation for both the deep learning and other conventional GPU workloads. 
This benefit comes with a very low area cost and only minimal adjustments of DRAM timing parameters to the standard HBM2 DRAM interface.
\end{abstract}
\cleardoublepage

% % 여기 수정할 것
\tableofcontents
\addcontentsline{toc}{chapter}{\contentsname}
\cleardoublepage

\listoftables
\addcontentsline{loc}{chapter}{\listtablename}
\cleardoublepage

\listoffigures
\addcontentsline{loc}{chapter}{\listfigurename}
\cleardoublepage

\pagenumbering{arabic}

\chapter{Introduction}
\input{ch1.tex}

\clearpage
\chapter{Background and Motivation}
\input{ch2.tex}

\chapter{Practical Partial Row Activation}
\input{ch3.tex}

\chapter{Evaluation}
\input{ch4.tex}

\chapter{Conclusion}
\input{ch5.tex}

\clearpage
%\section{Table insertion}
%\subsection{Basic}
%\url{https://www.tablesgenerator.com/latex_tables} creates the most basic template!
%
%\clearpage
%\subsection{Advanced}
%The most annoying part of latex work is table insertion.
%Size modification, highlights, and annotation in table are very annoying compared to Hancom, so many people turn to it.
%However, latex will feel much easier if you learn only the following introductions for graduation thesis.
%
%\textit{adjustbox}: It adjusts the overall size of the table. If not, a table may be generated  beyond the document scope.
%
%\textit{columncolor}: It shades the entire column to emphasize the results of my model.
%
%\textit{footnotemark and footnotetext}: Inside the table, the $\setminus$footnote does not work. 
%For this reason, footnotemark and footnotetext are used.
%Note that the page on which the annotation exists is not always the same (...) as the table.
%When the footnotemark is used several times inside the table, the annotation number becomes strange.
%In this case, use \textit{addtocounter}.
%\newline
%\newline
%If you need more than this, let's do googling.

\clearpage

%\addtocounter{footnote}{-1}
%\footnotetext{For a fair comparison, we report EntNet's result~\mycite{henaff2016tracking} which was jointly trained on all tasks. It was written in the appendix of the paper.}
%\addtocounter{footnote}{+1}
%\footnotetext{Our implementation. The result is different from what \myauthor{santoro2017simple} mentioned, which is caused by the initialization~\mycite{santoro2017simple}.}
%

%\chapter{Reference insertion}
%\section{Insert reference right before the period mark}
%Do not use conventional $\sim\setminus$cite\{\}.

%ex: blah blah~\cite{bishop2006}.
%Because of ( ), I defined `mycite' to appear reference number with [ ].

%ex: blah blah~\mycite{bishop2006}.


%\clearpage
%\section{Insert reference to the author name}
%Sometimes you need to put a reference to the author name. 
%To do this, I defined `myauthor'.
%ex: \myauthor{bishop2006} said blah blah~\mycite{bishop2006}.

\begin{bibpage}
	\bibliographystyle{ieeetr}
	\bibliography{references}
\end{bibpage}


\appendix
%\chapter{Appendix title 1}
%The appendix does not appear in the table of contents.
%In this case, open a .toc file (or click the table of contents page with holding ctrl) and add a line below the bibliography.
%
%\contentsline {chapter}{Bibliography}{8}{section*.9} % 현재 Bibliography
%\contentsline {chapter}{Appendix}{9}{section*.10} % {9}는 페이지 번호이므로 알아서 맞추자
%\contentsline {section}{\numberline {A}Appendix title 1}{9}{section.Alph0.1}
%\contentsline {section}{\numberline {B}Appendix title 2}{10}{section.Alph0.2}
%
%\chapter{Appendix title 2}

\keywordalt{CNN, GPU, DRAM 아키텍처, 에너지 효율, HBM2}
\begin{abstractalt}
GPU는 딥 러닝 애플리케이션을 실행하는 데 널리 사용된다.
오늘날의 high-end GPU는 HBM (High-Bandwidth Memory)과 같은 3D 적층 DRAM 기술을 채택하여 
엄청난 대역폭을 제공하므로 많은 전력을 소비한다.
GPU에서 수천 개의 동시 스레드가 발생하면 빈번한 row buffer conflict로 인해 상당한 양의 DRAM 에너지가 낭비된다.
이러한 낭비를 줄이기 위해 3D 적층 DRAM에 대한 partial row activation 기법을 제안한다.
풍부한 memory-level parallelism 이 있는 딥 러닝 워크 로드의 latency tolerance를 활용해서, DRAM latency를 지불하고 에너지 절감을 얻을 수 있다.
본 제안에서 딥 러닝 및 기타 기존 GPU 워크 로드에서 성능 저하를 최소화하면서 DRAM activation energy의 상당한 절감 효과를 보여준다.
본 제안은 매우 낮은 면적 비용으로 표준 HBM2 DRAM 인터페이스에 대한 DRAM 타이밍의 최소한의 변경만으로 구현할 수 있다는 장점이 있다.
\end{abstractalt}

\acknowledgement
\input{ack.tex}
%Thanks!

\end{document}