서브메뉴
검색
Architecture, Modeling, and Optimization of Photonic Neural Network Accelerators
Architecture, Modeling, and Optimization of Photonic Neural Network Accelerators
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211152834
- ISBN
- 9798384091998
- DDC
- 621.3
- 저자명
- Li, Shurui.
- 서명/저자
- Architecture, Modeling, and Optimization of Photonic Neural Network Accelerators
- 발행사항
- [Sl] : University of California, Los Angeles, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 255 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-03, Section: B.
- 주기사항
- Advisor: Gupta, Puneet.
- 학위논문주기
- Thesis (Ph.D.)--University of California, Los Angeles, 2024.
- 초록/해제
- 요약As artificial intelligence (AI) continues to advance, so too must the computational architectures that support it. Traditional Complementary Metal-Oxide-Semiconductor (CMOS)- based accelerators, while having served as the backbone of computing for decades, are now encountering significant limitations due to the slowing of Moore's Law. Photonic neural network accelerators emerge as a promising alternative, offering high-speed, parallel optical computations that can significantly enhance performance and energy efficiency.Despite their potential, photonic computing systems face significant challenges, particularly related to the conversion overhead from digital to analog signals and vice versa. The conversion overhead can drastically reduce the energy efficiency of photonic neural network accelerators. This dissertation addresses these challenges by introducing cross-layer optimizations and co-design strategies that span computing methods, circuits, architectures, and algorithms. Specifically, we focus on reducing the conversion overhead through innovative design and optimization techniques.Part 1: We introduce an innovative use of free-space optical systems, specifically 4F systems, to accelerate CNNs. By leveraging Fourier optics, we reduce the complexity of convolution operations from O(N2 ) to O(N), a feat unachievable by traditional electronic systems. This part includes the design, construction, and optimization of free-space optical CNN accelerators, demonstrating significant performance improvements and energy efficiency through experimental evaluations on datasets such as MNIST and CIFAR-10.Part 2: We delve into on-chip photonic neural network accelerators, presenting two pioneering architectures: PhotoFourier and ReFOCUS. PhotoFourier leverages the Joint Transform Correlator (JTC) approach to perform convolutions with reduced complexity and fewer photonic components. ReFOCUS builds upon this with innovative features like optical buffers and wavelength-division multiplexing (WDM), further improving energy and area efficiency. We demonstrate that these on-chip designs outperform contemporary photonic and CMOS accelerators in terms of throughput, power efficiency, and energy-delay product (EDP).Part 3: We focus on algorithmic innovations and theoretical analysis to enhance the efficiency of photonic and analog computing systems. We propose a weight pool compression algorithm that reduces storage requirements and memory traffic, enabling efficient deployment of large neural networks. This compression algorithm also has a promising synergy with analog and photonic neural network accelerators, which could avoid or drastically reduce the conversion overhead of weights. Additionally, since the ADC power is heavily dependent on the bitwidth (ADC resolution), we develop a comprehensive analytical model for partial sum precision requirements, optimizing the trade-offs between accuracy and energy efficiency in analog neural network accelerators.Experimental JTC challenges and findings: We also included a chapter dedicated to the challenges and non-idealities we observed in our JTC hardware prototype. We further provide sensitivity analysis for various non-linearity of the photodetectors. The goal of this section is to complement the architecture work with some experimental analysis and findings, and provide insights and directions for future work.Through our contributions, we advance the field of photonic neural network accelerators, providing new architectures, modeling techniques, and optimization strategies. Our findings demonstrate the potential of photonic technologies to achieve high-performance, energy-efficient AI computations, thereby addressing critical challenges in modern computing hardware and paving the way for future advancements in this rapidly evolving domain.
- 일반주제명
- Computer engineering
- 일반주제명
- Condensed matter physics
- 일반주제명
- Computer science
- 일반주제명
- Information technology
- 기타저자
- University of California, Los Angeles Electrical and Computer Engineering 0333
- 기본자료저록
- Dissertations Abstracts International. 86-03B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017164119
■00520250211152834
■006m o d
■007cr#unu||||||||
■020 ▼a9798384091998
■035 ▼a(MiAaPQ)AAI31561077
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a621.3
■1001 ▼aLi, Shurui.
■24510▼aArchitecture, Modeling, and Optimization of Photonic Neural Network Accelerators
■260 ▼a[Sl]▼bUniversity of California, Los Angeles▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a255 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-03, Section: B.
■500 ▼aAdvisor: Gupta, Puneet.
■5021 ▼aThesis (Ph.D.)--University of California, Los Angeles, 2024.
■520 ▼aAs artificial intelligence (AI) continues to advance, so too must the computational architectures that support it. Traditional Complementary Metal-Oxide-Semiconductor (CMOS)- based accelerators, while having served as the backbone of computing for decades, are now encountering significant limitations due to the slowing of Moore's Law. Photonic neural network accelerators emerge as a promising alternative, offering high-speed, parallel optical computations that can significantly enhance performance and energy efficiency.Despite their potential, photonic computing systems face significant challenges, particularly related to the conversion overhead from digital to analog signals and vice versa. The conversion overhead can drastically reduce the energy efficiency of photonic neural network accelerators. This dissertation addresses these challenges by introducing cross-layer optimizations and co-design strategies that span computing methods, circuits, architectures, and algorithms. Specifically, we focus on reducing the conversion overhead through innovative design and optimization techniques.Part 1: We introduce an innovative use of free-space optical systems, specifically 4F systems, to accelerate CNNs. By leveraging Fourier optics, we reduce the complexity of convolution operations from O(N2 ) to O(N), a feat unachievable by traditional electronic systems. This part includes the design, construction, and optimization of free-space optical CNN accelerators, demonstrating significant performance improvements and energy efficiency through experimental evaluations on datasets such as MNIST and CIFAR-10.Part 2: We delve into on-chip photonic neural network accelerators, presenting two pioneering architectures: PhotoFourier and ReFOCUS. PhotoFourier leverages the Joint Transform Correlator (JTC) approach to perform convolutions with reduced complexity and fewer photonic components. ReFOCUS builds upon this with innovative features like optical buffers and wavelength-division multiplexing (WDM), further improving energy and area efficiency. We demonstrate that these on-chip designs outperform contemporary photonic and CMOS accelerators in terms of throughput, power efficiency, and energy-delay product (EDP).Part 3: We focus on algorithmic innovations and theoretical analysis to enhance the efficiency of photonic and analog computing systems. We propose a weight pool compression algorithm that reduces storage requirements and memory traffic, enabling efficient deployment of large neural networks. This compression algorithm also has a promising synergy with analog and photonic neural network accelerators, which could avoid or drastically reduce the conversion overhead of weights. Additionally, since the ADC power is heavily dependent on the bitwidth (ADC resolution), we develop a comprehensive analytical model for partial sum precision requirements, optimizing the trade-offs between accuracy and energy efficiency in analog neural network accelerators.Experimental JTC challenges and findings: We also included a chapter dedicated to the challenges and non-idealities we observed in our JTC hardware prototype. We further provide sensitivity analysis for various non-linearity of the photodetectors. The goal of this section is to complement the architecture work with some experimental analysis and findings, and provide insights and directions for future work.Through our contributions, we advance the field of photonic neural network accelerators, providing new architectures, modeling techniques, and optimization strategies. Our findings demonstrate the potential of photonic technologies to achieve high-performance, energy-efficient AI computations, thereby addressing critical challenges in modern computing hardware and paving the way for future advancements in this rapidly evolving domain.
■590 ▼aSchool code: 0031.
■650 4▼aComputer engineering
■650 4▼aCondensed matter physics
■650 4▼aComputer science
■650 4▼aInformation technology
■653 ▼aComputer architecture
■653 ▼aNeural network compression
■653 ▼aOptical neural network
■653 ▼aPhotonic neural network accelerators
■653 ▼aEnergy-delay product
■690 ▼a0464
■690 ▼a0489
■690 ▼a0984
■690 ▼a0800
■690 ▼a0611
■71020▼aUniversity of California, Los Angeles▼bElectrical and Computer Engineering 0333.
■7730 ▼tDissertations Abstracts International▼g86-03B.
■790 ▼a0031
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17164119▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


