서브메뉴
검색
Vertical Integration in Open-Source Hardware-Software Co-Design for Accelerator Architecture Research
Vertical Integration in Open-Source Hardware-Software Co-Design for Accelerator Architecture Research
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105554
- ISBN
- 9798265403407
- DDC
- 570
- 서명/저자
- Vertical Integration in Open-Source Hardware-Software Co-Design for Accelerator Architecture Research
- 발행사항
- [Sl] : Georgia Institute of Technology, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 182 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-06, Section: B.
- 주기사항
- Advisor: Kim, Hyesoon.
- 학위논문주기
- Thesis (Ph.D.)--Georgia Institute of Technology, 2024.
- 초록/해제
- 요약The emergence of data-parallel architectures and General-Purpose Graphics Processing Unit (GPGPU) has enabled new opportunities to address the power limitations and scalability of multi-core processors, allowing for new ways to exploit the abundant data parallelism present in emerging big-data parallel applications such as machine learning and graph analytics. GPGPUs in particular, with their Single Instruction Multiple Threads (SIMT) execution model, heavily leverage data-parallel multi-threading to maximize throughput at a relatively low energy cost, leading the current race for energy efficiency (Green500) and application support with their accelerator-centric parallel programming models.For the past decade, GPGPU architecture research has mainly focused on cycle-level simulations that model the hardware architecture at the Intermediate Language (IL) level (PTX, HSAIL) because of the lack of open-source hardware implementation. Simulating complex hardware at the IL level can obfuscate several aspects of the micro-architecture that have a substantial impact on performance. The recent introduction of full-system Instruction Set Architecture (ISA)-based Graphics Processing Unit (GPU) model simulations has closed the evaluation gap with actual hardware but still remains limited as it does not cover other important areas such as run-time evaluation, power efficiency, reliability, and detailed microarchitecture evaluation that can be pursued when using Register Transfer Level (RTL)-level implementation. Several implementations of open-source GPGPU hardware have been proposed that provide a detailed micro-architectural description of various GPGPU's components. However, these implementations lack a detailed description of the cache subsystem and command processor, which are performance-critical components in the GPGPU. Also, the ISA used in those implementations is custom or proprietary, restricting application support and wide adoption.Additionally, graphics acceleration, one of the primary applications of GPUs and a driving force for most of the technological innovations in GPU architecture, has not received as much attention by the academic research community mostly due to the lack of an open ISA to enable microarchitecture explorations. The current area of graphics rendering moving to the cloud presents new challenges for graphics computation, including real-time latency, scalability, and hardware security, raising the importance to expand the study in this space.Two recent technological trends provide an opportunity to revisit and expand open-source GPGPUs for hardware research today:(1) The emergence of high-end Field-Programmable Gate Array (FPGA)s in the consumer market. Today's high-capacity FPGAs with floating-point Digital Signal Processor (DSP)s and large memory provide high computational capability at a lower energy budget that makes implementing a full-feature GPGPU with a detailed cache subsystem operating at a reasonable speed a possibility. (2) The advent of RISC-V with its free, open, and extensible ISA, provides a new level of freedom in designing hardware architectures at a lower cost that leverages its rich ecosystem of open-source software and compiler tools. Adopting the RISC-V ISA for a GPGPU processor architecture presents a solid base for wide-range adoption.However, hardware development presents two key challenges which have played an important role in limiting the development of open-source hardware in the past decade. The first challenge is the steep learning curve and expertise required to use hardware programming languages and development tools. The second challenge involves the limitations of simulation and debugging tools for hardware, RTL simulations are two to three orders of magnitude slower than cycle-level simulation, making it impractical for working on large-scale designs like GPGUs.
- 일반주제명
- Taxonomy
- 일반주제명
- 3-D graphics
- 일반주제명
- Vortices
- 일반주제명
- Bandwidths
- 일반주제명
- Computer engineering
- 기본자료저록
- Dissertations Abstracts International. 87-06B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2024 us c eng d■001000017360603
■00520260202105554
■006m o d
■007cr#unu||||||||
■020 ▼a9798265403407
■035 ▼a(MiAaPQ)AAI32315843
■035 ▼a(MiAaPQ)GeorgiaTech75184
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a570
■1001 ▼aTine, Blaise-Pascal.
■24510▼aVertical Integration in Open-Source Hardware-Software Co-Design for Accelerator Architecture Research
■260 ▼a[Sl]▼bGeorgia Institute of Technology▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a182 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-06, Section: B.
■500 ▼aAdvisor: Kim, Hyesoon.
■5021 ▼aThesis (Ph.D.)--Georgia Institute of Technology, 2024.
■520 ▼aThe emergence of data-parallel architectures and General-Purpose Graphics Processing Unit (GPGPU) has enabled new opportunities to address the power limitations and scalability of multi-core processors, allowing for new ways to exploit the abundant data parallelism present in emerging big-data parallel applications such as machine learning and graph analytics. GPGPUs in particular, with their Single Instruction Multiple Threads (SIMT) execution model, heavily leverage data-parallel multi-threading to maximize throughput at a relatively low energy cost, leading the current race for energy efficiency (Green500) and application support with their accelerator-centric parallel programming models.For the past decade, GPGPU architecture research has mainly focused on cycle-level simulations that model the hardware architecture at the Intermediate Language (IL) level (PTX, HSAIL) because of the lack of open-source hardware implementation. Simulating complex hardware at the IL level can obfuscate several aspects of the micro-architecture that have a substantial impact on performance. The recent introduction of full-system Instruction Set Architecture (ISA)-based Graphics Processing Unit (GPU) model simulations has closed the evaluation gap with actual hardware but still remains limited as it does not cover other important areas such as run-time evaluation, power efficiency, reliability, and detailed microarchitecture evaluation that can be pursued when using Register Transfer Level (RTL)-level implementation. Several implementations of open-source GPGPU hardware have been proposed that provide a detailed micro-architectural description of various GPGPU's components. However, these implementations lack a detailed description of the cache subsystem and command processor, which are performance-critical components in the GPGPU. Also, the ISA used in those implementations is custom or proprietary, restricting application support and wide adoption.Additionally, graphics acceleration, one of the primary applications of GPUs and a driving force for most of the technological innovations in GPU architecture, has not received as much attention by the academic research community mostly due to the lack of an open ISA to enable microarchitecture explorations. The current area of graphics rendering moving to the cloud presents new challenges for graphics computation, including real-time latency, scalability, and hardware security, raising the importance to expand the study in this space.Two recent technological trends provide an opportunity to revisit and expand open-source GPGPUs for hardware research today:(1) The emergence of high-end Field-Programmable Gate Array (FPGA)s in the consumer market. Today's high-capacity FPGAs with floating-point Digital Signal Processor (DSP)s and large memory provide high computational capability at a lower energy budget that makes implementing a full-feature GPGPU with a detailed cache subsystem operating at a reasonable speed a possibility. (2) The advent of RISC-V with its free, open, and extensible ISA, provides a new level of freedom in designing hardware architectures at a lower cost that leverages its rich ecosystem of open-source software and compiler tools. Adopting the RISC-V ISA for a GPGPU processor architecture presents a solid base for wide-range adoption.However, hardware development presents two key challenges which have played an important role in limiting the development of open-source hardware in the past decade. The first challenge is the steep learning curve and expertise required to use hardware programming languages and development tools. The second challenge involves the limitations of simulation and debugging tools for hardware, RTL simulations are two to three orders of magnitude slower than cycle-level simulation, making it impractical for working on large-scale designs like GPGUs.
■590 ▼aSchool code: 0078.
■650 4▼aTaxonomy
■650 4▼a3-D graphics
■650 4▼aVortices
■650 4▼aBandwidths
■650 4▼aComputer engineering
■653 ▼aGeneral-Purpose Graphics Processing Unit
■653 ▼aSingle Instruction Multiple Threads
■690 ▼a0464
■71020▼aGeorgia Institute of Technology.
■7730 ▼tDissertations Abstracts International▼g87-06B.
■790 ▼a0078
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17360603▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


