서브메뉴
검색
Overcoming Device Level Challenges to Enable Compute In-Memory with Emerging Non-Volatile Memory
Overcoming Device Level Challenges to Enable Compute In-Memory with Emerging Non-Volatile Memory
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105539
- ISBN
- 9798265400642
- DDC
- 620
- 저자명
- Crafton, Brian.
- 서명/저자
- Overcoming Device Level Challenges to Enable Compute In-Memory with Emerging Non-Volatile Memory
- 발행사항
- [Sl] : Georgia Institute of Technology, 2023
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2023
- 형태사항
- 145 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-05, Section: B.
- 주기사항
- Advisor: Raychowdhury, Arijit.
- 학위논문주기
- Thesis (Ph.D.)--Georgia Institute of Technology, 2023.
- 초록/해제
- 요약The ever growing performance gap between main memory and logic coupled with trends towards data-centric applications poses a significant challenge for modern computing systems. These applications demand higher memory capacity and bandwidth, while also lacking significant data re-use that on-chip SRAM cache has historically exploited to reduce bandwidth requirements. To further this challenge, applications of Machine Learning (ML) are trending towards more memory intensive operations [1] on smaller, resource constrained devices [2]. These challenges have motivated widespread adoption of hardware accelerators and High Bandwidth Memory (HBM) to maximize the compute throughput under modern memory constraints. Despite strong improvements, we face limitations that have inspired new memory technologies and techniques to enable future workloads on future computing systems.Compute In-Memory (CIM) is one such research thread that reads and accumulates multiple memory cells onto the same bit-line (BL). This increases memory bandwidth and performs (binary) multiplication and addition without the use of Complementary Metal-Oxide Semiconductor (CMOS) logic. At the same time, various Embedded Non-Volatile Memory (eNVM) such as Resistive Random Access Memory (RRAM), Phase Change Ran-dom Access Memory (PCRAM), Spin-Transfer-Torque Magnetic Random Access Memory (STT-MRAM), and Ferroeletric FET (FeFET) are being actively developed to enable high density non-volatile storage while being both logic and process compatible [3].The objective of the proposed research is to overcome challenges faced by CIM and eNVM. The combination of eNVM and CIM can greatly increase memory bandwidth and memory density, perform computation on the bitline of memory sub-arrays, and minimize data transport. However, CIM with eNVM introduces problems not before faced by tra-ditional CMOS and SRAM based designs. These challenges include device variation, IR-drop, high write power, and low endurance. In this thesis, we address several of these device-level challenges at the circuit and system level.In chapter 1 and chapter 2, we review relevant motivation and background, outline prior work, and discuss outstanding challenges. The first challenge we study is device variation and other error sources in CIM. In chapter 3, we characterize the impact of device variation by calibrating models based on measured data with foundry RRAM arrays. Next, propose a new algorithm based on device variation to increase both performance and accuracy for CIM designs. In chapter 5 and chapter 6, we continue our efforts on minimizing error in CIM. We extend our efforts to include both hard and soft faults in CIM. However, we look towards ways to actually reduce Bit Error Rate (BER), or the actual number of errors that occur. To do so, propose a new class of error correcting codes (ECC) for hard and soft errors in CIM. In chapter 4, we study the impact of high write latency, high write power, and low endurance have on Deep Neural Network (DNN) implementations using eNVM and CIM. Next, we outline the ways current work overcomes these challenges and highlight a key problem. Then we propose a new allocation algorithm and data flow based on input data distributions to maximize utilization and performance for compute-in memory based designs. Lastly, in chapter 7 we again extend our efforts on minimizing error in CIM to IR-drop. We first we characterize the impact of IR-drop on our existing designs and then evaluate different approaches to write verify. Using various voltages and pulse widths we program cells to offset IR-drop and demonstrate a 136.4x reduction in BER during CIM.
- 일반주제명
- Random access memory
- 일반주제명
- Semiconductors
- 일반주제명
- Bandwidths
- 일반주제명
- Neural networks
- 일반주제명
- CMOS
- 일반주제명
- Breakdowns
- 일반주제명
- Computer science
- 일반주제명
- Electrical engineering
- 기본자료저록
- Dissertations Abstracts International. 87-05B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2023 us c eng d■001000017360517
■00520260202105539
■006m o d
■007cr#unu||||||||
■020 ▼a9798265400642
■035 ▼a(MiAaPQ)AAI32315078
■035 ▼a(MiAaPQ)GeorgiaTech71997
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a620
■1001 ▼aCrafton, Brian.
■24510▼aOvercoming Device Level Challenges to Enable Compute In-Memory with Emerging Non-Volatile Memory
■260 ▼a[Sl]▼bGeorgia Institute of Technology▼c2023
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2023
■300 ▼a145 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-05, Section: B.
■500 ▼aAdvisor: Raychowdhury, Arijit.
■5021 ▼aThesis (Ph.D.)--Georgia Institute of Technology, 2023.
■520 ▼aThe ever growing performance gap between main memory and logic coupled with trends towards data-centric applications poses a significant challenge for modern computing systems. These applications demand higher memory capacity and bandwidth, while also lacking significant data re-use that on-chip SRAM cache has historically exploited to reduce bandwidth requirements. To further this challenge, applications of Machine Learning (ML) are trending towards more memory intensive operations [1] on smaller, resource constrained devices [2]. These challenges have motivated widespread adoption of hardware accelerators and High Bandwidth Memory (HBM) to maximize the compute throughput under modern memory constraints. Despite strong improvements, we face limitations that have inspired new memory technologies and techniques to enable future workloads on future computing systems.Compute In-Memory (CIM) is one such research thread that reads and accumulates multiple memory cells onto the same bit-line (BL). This increases memory bandwidth and performs (binary) multiplication and addition without the use of Complementary Metal-Oxide Semiconductor (CMOS) logic. At the same time, various Embedded Non-Volatile Memory (eNVM) such as Resistive Random Access Memory (RRAM), Phase Change Ran-dom Access Memory (PCRAM), Spin-Transfer-Torque Magnetic Random Access Memory (STT-MRAM), and Ferroeletric FET (FeFET) are being actively developed to enable high density non-volatile storage while being both logic and process compatible [3].The objective of the proposed research is to overcome challenges faced by CIM and eNVM. The combination of eNVM and CIM can greatly increase memory bandwidth and memory density, perform computation on the bitline of memory sub-arrays, and minimize data transport. However, CIM with eNVM introduces problems not before faced by tra-ditional CMOS and SRAM based designs. These challenges include device variation, IR-drop, high write power, and low endurance. In this thesis, we address several of these device-level challenges at the circuit and system level.In chapter 1 and chapter 2, we review relevant motivation and background, outline prior work, and discuss outstanding challenges. The first challenge we study is device variation and other error sources in CIM. In chapter 3, we characterize the impact of device variation by calibrating models based on measured data with foundry RRAM arrays. Next, propose a new algorithm based on device variation to increase both performance and accuracy for CIM designs. In chapter 5 and chapter 6, we continue our efforts on minimizing error in CIM. We extend our efforts to include both hard and soft faults in CIM. However, we look towards ways to actually reduce Bit Error Rate (BER), or the actual number of errors that occur. To do so, propose a new class of error correcting codes (ECC) for hard and soft errors in CIM. In chapter 4, we study the impact of high write latency, high write power, and low endurance have on Deep Neural Network (DNN) implementations using eNVM and CIM. Next, we outline the ways current work overcomes these challenges and highlight a key problem. Then we propose a new allocation algorithm and data flow based on input data distributions to maximize utilization and performance for compute-in memory based designs. Lastly, in chapter 7 we again extend our efforts on minimizing error in CIM to IR-drop. We first we characterize the impact of IR-drop on our existing designs and then evaluate different approaches to write verify. Using various voltages and pulse widths we program cells to offset IR-drop and demonstrate a 136.4x reduction in BER during CIM.
■590 ▼aSchool code: 0078.
■650 4▼aRandom access memory
■650 4▼aSemiconductors
■650 4▼aError correction & detection
■650 4▼aBandwidths
■650 4▼aNeural networks
■650 4▼aCMOS
■650 4▼aBreakdowns
■650 4▼aComputer science
■650 4▼aElectrical engineering
■690 ▼a0800
■690 ▼a0984
■690 ▼a0544
■71020▼aGeorgia Institute of Technology.
■7730 ▼tDissertations Abstracts International▼g87-05B.
■790 ▼a0078
■791 ▼aPh.D.
■792 ▼a2023
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17360517▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


