본문

서브메뉴

Applications of Machine Learning and Optimization Techniques to Solve Modern Data Movement Challenges
Applications of Machine Learning and Optimization Techniques to Solve Modern Data Movement...
Applications of Machine Learning and Optimization Techniques to Solve Modern Data Movement Challenges

상세정보

자료유형  
 학위논문 서양
최종처리일시  
20260202104854
ISBN  
9798288816901
DDC  
620
저자명  
Sudhakaran, Sunil.
서명/저자  
Applications of Machine Learning and Optimization Techniques to Solve Modern Data Movement Challenges
발행사항  
[Sl] : Stanford University, 2025
발행사항  
Ann Arbor : ProQuest Dissertations & Theses, 2025
형태사항  
129 p
주기사항  
Source: Dissertations Abstracts International, Volume: 87-02, Section: B.
주기사항  
Advisor: Horowitz, Mark.
학위논문주기  
Thesis (Ph.D.)--Stanford University, 2025.
초록/해제  
요약Artificial-intelligence (AI) models are growing rapidly, moving the primary system bottleneck from core compute to the high-speed links that transfer data among accelerators, memory, and processors. As data rates climb, frequency-dependent loss, crosstalk, and supply noise become more formidable, yet the I/O power budget must stay tight so as not to encroach on compute resources. Furthermore, because these impairments drift with temperature and workloads, transceivers must periodically readjust their settings. We tackle this problem by combining formal optimization with lightweight machine-learning techniques. First, we propose a generalized decision-feedback equalizer that employs multiple slicer levels and a programmable look-up table, treating inter-symbol interference and crosstalk as observable noise sources. Finding the optimal slicer levels and table entries is a discrete optimization problem solved by integer-linear programming (ILP) methods. However, ILP is too memory-intensive for on-line use. Instead, we train a compact convolutional neural network supervised by ILP-derived solutions and deploy it on a 1GHz RISC-V microcontroller, where a forward pass runs in under 5ms and cuts memory demand from gigabytes to kilobytes. As a complementary measure, we explore constrained coding to mitigate link impairments. Encoder and decoder synthesis is first framed as a satisfiability-modulo-theories (SMT) problem, reducing solution time from months to days for codes up to eight bits. To scale further, we recast the codeword assignment problem as a bipartite-matching task and solve it with a reinforcement-learning agent. This approach handles codes up to ten bits and discovers assignments that minimize gate count, keeping the I/O footprint small and preserving silicon area for compute. Together, these contributions increase link performance without adversely increasing power and area requirements, charting a practical path toward faster and more energy-efficient data movement in next-generation AI systems.
일반주제명  
Silicon
일반주제명  
Transmitters
일반주제명  
Photonics
일반주제명  
Codes
일반주제명  
Optimization techniques
일반주제명  
Parameter estimation
일반주제명  
Computer engineering
일반주제명  
Electrical engineering
키워드  
Machine-learning techniques
키워드  
Integer-linear programming methods
키워드  
Encoder and decoder synthesis
기타저자  
Stanford University.
기본자료저록  
Dissertations Abstracts International. 87-02B.
전자적 위치 및 접속  
로그인 후 원문을 볼 수 있습니다.

MARC

 008260126s2025        us                              c    eng  d
■001000017359239
■00520260202104854
■006m          o    d                
■007cr#unu||||||||
■020    ▼a9798288816901
■035    ▼a(MiAaPQ)AAI32200991
■035    ▼a(MiAaPQ)Stanfordsg461wr4702
■040    ▼aMiAaPQ▼cMiAaPQ
■0820  ▼a620
■1001  ▼aSudhakaran,  Sunil.
■24510▼aApplications  of  Machine  Learning  and  Optimization  Techniques  to  Solve  Modern  Data  Movement  Challenges
■260    ▼a[Sl]▼bStanford  University▼c2025
■260  1▼aAnn  Arbor▼bProQuest  Dissertations  &  Theses▼c2025
■300    ▼a129  p
■500    ▼aSource:  Dissertations  Abstracts  International,  Volume:  87-02,  Section:  B.
■500    ▼aAdvisor:  Horowitz,  Mark.
■5021  ▼aThesis  (Ph.D.)--Stanford  University,  2025.
■520    ▼aArtificial-intelligence  (AI)  models  are  growing  rapidly,  moving  the  primary  system  bottleneck  from  core  compute  to  the  high-speed  links  that  transfer  data  among  accelerators,  memory,  and  processors.  As  data  rates  climb,  frequency-dependent  loss,  crosstalk,  and  supply  noise  become  more  formidable,  yet  the  I/O  power  budget  must  stay  tight  so  as  not  to  encroach  on  compute  resources.  Furthermore,  because  these  impairments  drift  with  temperature  and  workloads,  transceivers  must  periodically  readjust  their  settings.  We  tackle  this  problem  by  combining  formal  optimization  with  lightweight  machine-learning  techniques.  First,  we  propose  a  generalized  decision-feedback  equalizer  that  employs  multiple  slicer  levels  and  a  programmable  look-up  table,  treating  inter-symbol  interference  and  crosstalk  as  observable  noise  sources.  Finding  the  optimal  slicer  levels  and  table  entries  is  a  discrete  optimization  problem  solved  by  integer-linear  programming  (ILP)  methods.  However,  ILP  is  too  memory-intensive  for  on-line  use.  Instead,  we  train  a  compact  convolutional  neural  network  supervised  by  ILP-derived  solutions  and  deploy  it  on  a  1GHz  RISC-V  microcontroller,  where  a  forward  pass  runs  in  under  5ms  and  cuts  memory  demand  from  gigabytes  to  kilobytes.  As  a  complementary  measure,  we  explore  constrained  coding  to  mitigate  link  impairments.  Encoder  and  decoder  synthesis  is  first  framed  as  a  satisfiability-modulo-theories  (SMT)  problem,  reducing  solution  time  from  months  to  days  for  codes  up  to  eight  bits.  To  scale  further,  we  recast  the  codeword  assignment  problem  as  a  bipartite-matching  task  and  solve  it  with  a  reinforcement-learning  agent.  This  approach  handles  codes  up  to  ten  bits  and  discovers  assignments  that  minimize  gate  count,  keeping  the  I/O  footprint  small  and  preserving  silicon  area  for  compute.  Together,  these  contributions  increase  link  performance  without  adversely  increasing  power  and  area  requirements,  charting  a  practical  path  toward  faster  and  more  energy-efficient  data  movement  in  next-generation  AI  systems.
■590    ▼aSchool  code:  0212.
■650  4▼aSilicon
■650  4▼aTransmitters
■650  4▼aPhotonics
■650  4▼aCodes
■650  4▼aOptimization  techniques
■650  4▼aParameter  estimation
■650  4▼aComputer  engineering
■650  4▼aElectrical  engineering
■653    ▼aMachine-learning  techniques
■653    ▼aInteger-linear  programming  methods
■653    ▼aEncoder  and  decoder  synthesis
■690    ▼a0544
■690    ▼a0464
■690    ▼a0800
■71020▼aStanford  University.
■7730  ▼tDissertations  Abstracts  International▼g87-02B.
■790    ▼a0212
■791    ▼aPh.D.
■792    ▼a2025
■793    ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17359239▼nKERIS▼z이  자료의  원문은  한국교육학술정보원에서  제공합니다.

미리보기

내보내기

chatGPT토론

Ai 추천 관련 도서


    신착도서 더보기
    최근 3년간 통계입니다.

    소장정보

    • 예약
    • 소재불명신고
    • 나의폴더
    • 우선정리요청
    • 비도서대출신청
    • 야간 도서대출신청
    소장자료
    등록번호 청구기호 소장처 대출가능여부 대출정보
    TF14910 전자도서 대출가능 마이폴더 부재도서신고 비도서대출신청 야간 도서대출신청

    * 대출중인 자료에 한하여 예약이 가능합니다. 예약을 원하시면 예약버튼을 클릭하십시오.

    해당 도서를 다른 이용자가 함께 대출한 도서

    관련 인기도서

    로그인 후 이용 가능합니다.