서브메뉴
검색
Architecture and Circuit Design Optimization for Compute-In-Memory
Architecture and Circuit Design Optimization for Compute-In-Memory
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105537
- ISBN
- 9798263392802
- DDC
- 620
- 저자명
- Jiang, Hongwu.
- 서명/저자
- Architecture and Circuit Design Optimization for Compute-In-Memory
- 발행사항
- [Sl] : Georgia Institute of Technology, 2022
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2022
- 형태사항
- 126 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-05, Section: A.
- 주기사항
- Advisor: Yu, Shimeng.
- 학위논문주기
- Thesis (Ph.D.)--Georgia Institute of Technology, 2022.
- 초록/해제
- 요약The main objective of this thesis is to optimize computing-in-memory (CIM) design for accelerating Deep Neural Network (DNN) algorithms. As compute peripheries such as analog-to-digital converters (ADCs) introduce significant overhead in CIM inference design, the first part of the research focuses on circuit optimizations for inmemory computing. In the first work, we comprehensively explore the tradeoffs involving different types of ADCs and investigate a new ADC design especially suited for the CIM, which performs the analog shift-add for multiple weight significance bits, improving the throughput and energy efficiency under similar area constraints. In the second work, we propose a resistive random access memory (RRAM) based ADC-free in-memory compute scheme validated with a prototype chip in TSMC 40nm process, which can significantly improve the hardware performance over the conventional CIM designs while achieving near-software classification accuracy on ImageNet and CIFAR-10/-100 dataset.In the second part of the thesis, the research focuses on hardware support for CIM on-chip training. To maximize hardware reuse of CIM weight stationary dataflow, we propose the CIM training architectures with the transpose weight mapping strategy. The cell design and periphery circuitry are modified to support bi-directional computing efficiently. A novel solution of signed number multiplication is also proposed to handle the negative inputs in backpropagation. Based on the silicon measurement data on a two-way SRAM-based prototype chip in TSMC 28nm process, we comprehensively explore the hardware performance for the entire SRAM-based architecture for DNN on-chip training.
- 일반주제명
- Silicon
- 일반주제명
- Design optimization
- 일반주제명
- Random access memory
- 일반주제명
- Deep learning
- 일반주제명
- Back propagation
- 일반주제명
- Macros
- 일반주제명
- Semiconductors
- 일반주제명
- Neural networks
- 일반주제명
- Circuits
- 일반주제명
- Energy efficiency
- 일반주제명
- CMOS
- 일반주제명
- Breakdowns
- 일반주제명
- Digitization
- 일반주제명
- Transistors
- 일반주제명
- Energy consumption
- 일반주제명
- Linear algebra
- 일반주제명
- Computer science
- 일반주제명
- Electrical engineering
- 일반주제명
- Sustainability
- 기본자료저록
- Dissertations Abstracts International. 87-05A.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2022 us c eng d■001000017360504
■00520260202105537
■006m o d
■007cr#unu||||||||
■020 ▼a9798263392802
■035 ▼a(MiAaPQ)AAI32314896
■035 ▼a(MiAaPQ)GeorgiaTech70137
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a620
■1001 ▼aJiang, Hongwu.
■24510▼aArchitecture and Circuit Design Optimization for Compute-In-Memory
■260 ▼a[Sl]▼bGeorgia Institute of Technology▼c2022
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2022
■300 ▼a126 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-05, Section: A.
■500 ▼aAdvisor: Yu, Shimeng.
■5021 ▼aThesis (Ph.D.)--Georgia Institute of Technology, 2022.
■520 ▼aThe main objective of this thesis is to optimize computing-in-memory (CIM) design for accelerating Deep Neural Network (DNN) algorithms. As compute peripheries such as analog-to-digital converters (ADCs) introduce significant overhead in CIM inference design, the first part of the research focuses on circuit optimizations for inmemory computing. In the first work, we comprehensively explore the tradeoffs involving different types of ADCs and investigate a new ADC design especially suited for the CIM, which performs the analog shift-add for multiple weight significance bits, improving the throughput and energy efficiency under similar area constraints. In the second work, we propose a resistive random access memory (RRAM) based ADC-free in-memory compute scheme validated with a prototype chip in TSMC 40nm process, which can significantly improve the hardware performance over the conventional CIM designs while achieving near-software classification accuracy on ImageNet and CIFAR-10/-100 dataset.In the second part of the thesis, the research focuses on hardware support for CIM on-chip training. To maximize hardware reuse of CIM weight stationary dataflow, we propose the CIM training architectures with the transpose weight mapping strategy. The cell design and periphery circuitry are modified to support bi-directional computing efficiently. A novel solution of signed number multiplication is also proposed to handle the negative inputs in backpropagation. Based on the silicon measurement data on a two-way SRAM-based prototype chip in TSMC 28nm process, we comprehensively explore the hardware performance for the entire SRAM-based architecture for DNN on-chip training.
■590 ▼aSchool code: 0078.
■650 4▼aSilicon
■650 4▼aDesign optimization
■650 4▼aRandom access memory
■650 4▼aDeep learning
■650 4▼aBack propagation
■650 4▼aMacros
■650 4▼aSemiconductors
■650 4▼aNeural networks
■650 4▼aCircuits
■650 4▼aEnergy efficiency
■650 4▼aCMOS
■650 4▼aBreakdowns
■650 4▼aDigitization
■650 4▼aTransistors
■650 4▼aEnergy consumption
■650 4▼aLinear algebra
■650 4▼aComputer science
■650 4▼aElectrical engineering
■650 4▼aSustainability
■690 ▼a0800
■690 ▼a0984
■690 ▼a0544
■690 ▼a0640
■71020▼aGeorgia Institute of Technology.
■7730 ▼tDissertations Abstracts International▼g87-05A.
■790 ▼a0078
■791 ▼aPh.D.
■792 ▼a2022
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17360504▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


