서브메뉴
검색
Methodologies for Modeling and Optimization of 2.5-D and 3-D Integration Architectures for Compute-in-Memory Applications
Methodologies for Modeling and Optimization of 2.5-D and 3-D Integration Architectures for Compute-in-Memory Applications
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260209102915
- ISBN
- 9798265403506
- DDC
- 790
- 저자명
- Kaul, Ankit.
- 서명/저자
- Methodologies for Modeling and Optimization of 2.5-D and 3-D Integration Architectures for Compute-in-Memory Applications
- 발행사항
- [Sl] : Georgia Institute of Technology, 2023
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2023
- 형태사항
- 160 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-06, Section: B.
- 주기사항
- Advisor: Bakir, Muhannad S.
- 학위논문주기
- Thesis (Ph.D.)--Georgia Institute of Technology, 2023.
- 초록/해제
- 요약In the wake of data-intensive computing, von Neumann-based traditional architectures and conventional methods of integration such as monolithic 2-D, are facing multiple challenges with reduced performance and higher costs. Such systems suffer from higher latency of communication between the memory hierarchy and processing elements, high power consumption, and increased hardware cost. To support the rising demand for emerging applications and data-intensive workloads, such systems require higher memory bandwidth to reduce latency and more efficient devices to improve energy-per-operation. Additionally, non-linear trends in device densities and energy per operation, and the rising design costs of advanced technology nodes have made conventional feature scaling an expensive pursuit.Due to the challenges with traditional von Neumann-based devices, new paradigms for compute, memory, communication, and integration have emerged. compute-in-memory (CIM) has been proposed as a potential paradigm for energy-efficient compute by reduced data movement and increased parallelism in image recognition and language translation computations. Further, a growing need for higher logic-memory bandwidth and lower chip-to-chip signal interconnection delay have led to a technological push towards heterogeneous integration. In this work, we propose methodologies to model physical effects and optimize design parameters in heterogeneous integration (HI) architectures for compute-in-memory hardware.First, the design trade-offs of including a power delivery network (PDN) and metal-insulator-metal (MIM) capacitors in bridge-chip based 2.5-D heterogeneous platforms are investigated. It is demonstrated that including the PDN (and MIM capacitors) in the bridge-chip can be an effective technique to reduce both DC-IR-drop and Ldi/dt noise. Next, to address the power delivery challenges in three dimensional heterogeneous integration (3-D-HI), a systematic technology and design space exploration of power delivery for 3-D-HI CIM systems is presented. A fast analysis flow facilitating early design-space exploration between power delivery design parameters and CIM performance metrics is proposed. By co-optimizing 3-D PDN and successive-approximation-register analog-to-digital converter (SAR-ADC) design parameters a balanced 3-D CIM design is demonstrated compared to a 3-D unoptimized implementation at iso-power and iso-area.Next, the thermal impact of different 3-D-HI architectures on the reliability of 3-D-integrated binary resistive random access memory (RRAM) devices for CIM applications is quantified. A device-integration reliability evaluation methodology is proposed that can be used to quantify the direct impact of integration design parameters on CIM inference accuracy. Using this flow, heterogeneous 3-D logic-memory CIM accelerator designs are benchmarked against monolithic 2-D and balanced integration design parameters for maximized 3-D CIM inference accuracy are reported. The benchmark framework is released as an open-source tool for the research community.A 3-D polylithic architecture is proposed that represents a densely integrated system divided into multiple device tiers where custom chiplets, such as power management IP, I/O drivers, and memory are embedded into the back-end of a base tier with extreme efficient signaling and large bandwidth density. Design optimization strategies for PDN in polylithic 3-D integration are presented. The scope includes a detailed design space exploration of the power supply noise effects in polylithic 3-D architectures. The thermal constraints for polylithic 3-D are evaluated with aggressive cooling to investigate thermal limits from transient- and steady-state perspectives.
- 일반주제명
- Design optimization
- 일반주제명
- Cooling
- 일반주제명
- Electricity generation
- 일반주제명
- Electric power
- 일반주제명
- Computer science
- 기본자료저록
- Dissertations Abstracts International. 87-06B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260203s2023 us c eng d■001000017366017
■00520260209102915
■006m o d
■007cr#unu||||||||
■020 ▼a9798265403506
■035 ▼a(MiAaPQ)AAI32316219
■035 ▼a(MiAaPQ)GeorgiaTech72787
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a790
■1001 ▼aKaul, Ankit.
■24510▼aMethodologies for Modeling and Optimization of 2.5-D and 3-D Integration Architectures for Compute-in-Memory Applications
■260 ▼a[Sl]▼bGeorgia Institute of Technology▼c2023
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2023
■300 ▼a160 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-06, Section: B.
■500 ▼aAdvisor: Bakir, Muhannad S.
■5021 ▼aThesis (Ph.D.)--Georgia Institute of Technology, 2023.
■520 ▼aIn the wake of data-intensive computing, von Neumann-based traditional architectures and conventional methods of integration such as monolithic 2-D, are facing multiple challenges with reduced performance and higher costs. Such systems suffer from higher latency of communication between the memory hierarchy and processing elements, high power consumption, and increased hardware cost. To support the rising demand for emerging applications and data-intensive workloads, such systems require higher memory bandwidth to reduce latency and more efficient devices to improve energy-per-operation. Additionally, non-linear trends in device densities and energy per operation, and the rising design costs of advanced technology nodes have made conventional feature scaling an expensive pursuit.Due to the challenges with traditional von Neumann-based devices, new paradigms for compute, memory, communication, and integration have emerged. compute-in-memory (CIM) has been proposed as a potential paradigm for energy-efficient compute by reduced data movement and increased parallelism in image recognition and language translation computations. Further, a growing need for higher logic-memory bandwidth and lower chip-to-chip signal interconnection delay have led to a technological push towards heterogeneous integration. In this work, we propose methodologies to model physical effects and optimize design parameters in heterogeneous integration (HI) architectures for compute-in-memory hardware.First, the design trade-offs of including a power delivery network (PDN) and metal-insulator-metal (MIM) capacitors in bridge-chip based 2.5-D heterogeneous platforms are investigated. It is demonstrated that including the PDN (and MIM capacitors) in the bridge-chip can be an effective technique to reduce both DC-IR-drop and Ldi/dt noise. Next, to address the power delivery challenges in three dimensional heterogeneous integration (3-D-HI), a systematic technology and design space exploration of power delivery for 3-D-HI CIM systems is presented. A fast analysis flow facilitating early design-space exploration between power delivery design parameters and CIM performance metrics is proposed. By co-optimizing 3-D PDN and successive-approximation-register analog-to-digital converter (SAR-ADC) design parameters a balanced 3-D CIM design is demonstrated compared to a 3-D unoptimized implementation at iso-power and iso-area.Next, the thermal impact of different 3-D-HI architectures on the reliability of 3-D-integrated binary resistive random access memory (RRAM) devices for CIM applications is quantified. A device-integration reliability evaluation methodology is proposed that can be used to quantify the direct impact of integration design parameters on CIM inference accuracy. Using this flow, heterogeneous 3-D logic-memory CIM accelerator designs are benchmarked against monolithic 2-D and balanced integration design parameters for maximized 3-D CIM inference accuracy are reported. The benchmark framework is released as an open-source tool for the research community.A 3-D polylithic architecture is proposed that represents a densely integrated system divided into multiple device tiers where custom chiplets, such as power management IP, I/O drivers, and memory are embedded into the back-end of a base tier with extreme efficient signaling and large bandwidth density. Design optimization strategies for PDN in polylithic 3-D integration are presented. The scope includes a detailed design space exploration of the power supply noise effects in polylithic 3-D architectures. The thermal constraints for polylithic 3-D are evaluated with aggressive cooling to investigate thermal limits from transient- and steady-state perspectives.
■590 ▼aSchool code: 0078.
■650 4▼aDesign optimization
■650 4▼aCooling
■650 4▼aElectricity generation
■650 4▼aElectric power
■650 4▼aComputer science
■653 ▼aCompute-in-memory
■653 ▼aDesign parameters
■690 ▼a0984
■71020▼aGeorgia Institute of Technology.
■7730 ▼tDissertations Abstracts International▼g87-06B.
■790 ▼a0078
■791 ▼aPh.D.
■792 ▼a2023
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17366017▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


