Jiwon Lee
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Jiwon Lee has a total of 43 co-authors. Here are the top co-authors by number of shared publications:
- Won Woo Ro
Yonsei University
7 shared publications - Jaewon Kwon
Yonsei University, Republic of Korea
3 shared publications - Minjae Kim
Department of Electrical and Electronic Engineering, Yonsei University, Seoul, South Korea
3 shared publications - Jiwan Kim
Embedded Systems and Computer Architecture Lab, Yonsei University, Seodaemun-gu, Seoul, Korea (the Republic of)
2 shared publications - Won Hur
LG Electronics, Seoul, South Korea
2 shared publications - Cheolhwan Kim
Department of Electrical and Electronic Engineering, Yonsei University, Seoul, South Korea
1 shared publication
Latest Publications
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computing
- Vol. 14 (3) , 1-14
SAPPhire: Practical Tiered Memory Systems with Memory Availability Prioritization
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computing
- Vol. 14 (3) , 1-14
SAPPhire: Practical Tiered Memory Systems with Memory Availability Prioritization
- Open Access
- 17 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 418-431
LATPC: Accelerating GPU Address Translation Using Locality-Aware TLB Prefetching and MSHR Compression
- 17 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 418-431
Open AccessLATPC: Accelerating GPU Address Translation Using Locality-Aware TLB Prefetching and MSHR Compression
- 14 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 24 (2) , 385-388
REDIT: Redirection-Enabled Memory-Side Directory Architecture for CXL Memory Fabric
The emergence of CXL memory fabrics enables composable, shared memory across multiple server hosts. While this significantly expands memory capacity, it also introduces cache coherence challenges between hosts sharing memory nodes. To address these, prior work explored three coherence control f low (CF) designs: host-side (HSD), switch-side (SWSD), and memory-side (MSD). Our analysis shows that MSD-CF is most effective, as it avoids the round-trip latency of HSD-CF and overcomes the scalability limits of SWSD-CF by distributing directories across memory devices. Using MSD-CF as our base line, we observe that it still experiences performance degradation when the directory reaches capacity, forcing fallback to host-side invalidation. To overcome this, we propose REDIT (Redirection Enabled memory-side DIrecTory architecture), which redirects overflow entries to underutilized directories and reclaims unused entries with minimal runtime cost. Evaluation shows that REDIT improves performance by 3.3× over HSD-CF, 2.0× over SWSD CF, and 1.6× over MSD-CF with only 1.5 KB of area overhead.
- 14 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 24 (2) , 385-388
REDIT: Redirection-Enabled Memory-Side Directory Architecture for CXL Memory Fabric
The emergence of CXL memory fabrics enables composable, shared memory across multiple server hosts. While this significantly expands memory capacity, it also introduces cache coherence challenges between hosts sharing memory nodes. To address these, prior work explored three coherence control f low (CF) designs: host-side (HSD), switch-side (SWSD), and memory-side (MSD). Our analysis shows that MSD-CF is most effective, as it avoids the round-trip latency of HSD-CF and overcomes the scalability limits of SWSD-CF by distributing directories across memory devices. Using MSD-CF as our base line, we observe that it still experiences performance degradation when the directory reaches capacity, forcing fallback to host-side invalidation. To overcome this, we propose REDIT (Redirection Enabled memory-side DIrecTory architecture), which redirects overflow entries to underutilized directories and reclaims unused entries with minimal runtime cost. Evaluation shows that REDIT improves performance by 3.3× over HSD-CF, 2.0× over SWSD CF, and 1.6× over MSD-CF with only 1.5 KB of area overhead.
- Open Access
- 20 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 122-136
Heliostat: Harnessing Ray Tracing Accelerators for Page Table Walks
- 20 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 122-136
Open AccessHeliostat: Harnessing Ray Tracing Accelerators for Page Table Walks
- 19 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 74 (6) , 1872-1885
HashScape: Leveraging Virtual Address Dynamics for Efficient Hashed Page Tables
The evolving memory landscape for larger capacity prompts alternative approaches due to scalability challenges in multi-level page tables, which require multiple serial memory accesses for address translation. Hashed Page Tables (HPTs) have gained attention for ideally facilitating a single memory access per translation. However, current HPTs increase minor page fault latency, thereby impeding its superiority over conventional multi-level page table design. This paper provides a comprehensive analysis of HPTs regarding minor page fault latency concerning memory management subsystems. In particular, we demonstrate how feasibility issues in memory management with HPTs can escalate minor page fault latency. We observe that different page types in HPTs (anon pages and page caches) exhibit distinct behaviors on the occurrence of minor page faults, indicating a significant correlation between page types and minor page faults. To address these challenges, we propose HashScape, a scheme that harmonizes with memory management using tailored HPTs per segment and size-tailored allocation via Virtual Memory Areas. Our evaluation demonstrates that HashScape significantly improves the insertion latency, with average, 95th, and 99th percentiles improving by 1.8$\boldsymbol{\times}$ , 1.9$\boldsymbol{\times}$ , and 2.2$\boldsymbol{\times}$ , respectively, resulting in an overall 10% reduction in minor page fault latency compared to a state-of-the-art HPT design.- 19 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 74 (6) , 1872-1885
HashScape: Leveraging Virtual Address Dynamics for Efficient Hashed Page Tables
The evolving memory landscape for larger capacity prompts alternative approaches due to scalability challenges in multi-level page tables, which require multiple serial memory accesses for address translation. Hashed Page Tables (HPTs) have gained attention for ideally facilitating a single memory access per translation. However, current HPTs increase minor page fault latency, thereby impeding its superiority over conventional multi-level page table design. This paper provides a comprehensive analysis of HPTs regarding minor page fault latency concerning memory management subsystems. In particular, we demonstrate how feasibility issues in memory management with HPTs can escalate minor page fault latency. We observe that different page types in HPTs (anon pages and page caches) exhibit distinct behaviors on the occurrence of minor page faults, indicating a significant correlation between page types and minor page faults. To address these challenges, we propose HashScape, a scheme that harmonizes with memory management using tailored HPTs per segment and size-tailored allocation via Virtual Memory Areas. Our evaluation demonstrates that HashScape significantly improves the insertion latency, with average, 95th, and 99th percentiles improving by 1.8$\boldsymbol{\times}$ , 1.9$\boldsymbol{\times}$ , and 2.2$\boldsymbol{\times}$ , respectively, resulting in an overall 10% reduction in minor page fault latency compared to a state-of-the-art HPT design.- Open Access
- 23 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Geneva: A Dynamic Confluence of Speculative Execution and In-Order Commitment Windows
Modern out-of-order microprocessors are increasingly expanding resources such as reorder buffer (ROB) and instruction queue (IQ) for memory-level parallelism (MLP). While this expansion effectively addresses the memory wall challenge, it also incurs notable cost and energy trade-offs. To tackle this, we propose Geneva, a microarchitecture that improves performance and saves energy. Geneva reallocates a portion of an ROB to serve as a dynamic queue (DQ), used as an ROB, IQ, or both depending on operational needs. Geneva saves energy by 15.6% and improves performance by 2.6% compared to the conventional out-of-order core.
- 23 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Open AccessGeneva: A Dynamic Confluence of Speculative Execution and In-Order Commitment Windows
Modern out-of-order microprocessors are increasingly expanding resources such as reorder buffer (ROB) and instruction queue (IQ) for memory-level parallelism (MLP). While this expansion effectively addresses the memory wall challenge, it also incurs notable cost and energy trade-offs. To tackle this, we propose Geneva, a microarchitecture that improves performance and saves energy. Geneva reallocates a portion of an ROB to serve as a dynamic queue (DQ), used as an ROB, IQ, or both depending on operational needs. Geneva saves energy by 15.6% and improves performance by 2.6% compared to the conventional out-of-order core.
- 25 June 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-5
Adaptive Data Prefetcher with Probability Learning in LLC
Data prefetching improves the performance of CPUs by predicting which data will be used in the near future. This paper introduces Adaptive Data Prefetcher (ADP), an effective prefetching technique that enhances the coverage of existing prefetchers. First, ADP consists of two prefetcher blocks. One is a path confidence-based lookahead prefetcher from prior work. The other is an LSTM prefetcher that predicts the next data address to be used based on block offset and page pattern history. The LSTM prefetcher is trained with the memory access patterns that the first prefetcher cannot predict, and thus it can further improve the performance of prefetching. Second, ADP adaptively selects a prefetcher which has a higher probability to hit the address of incoming memory accesses. Together, these features improve the coverage and accuracy of prefetching. In our analysis, we find that ADP achieves an IPC improvement of 24.2% over the confidence-based lookahead prefetcher.
- 25 June 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-5
Adaptive Data Prefetcher with Probability Learning in LLC
Data prefetching improves the performance of CPUs by predicting which data will be used in the near future. This paper introduces Adaptive Data Prefetcher (ADP), an effective prefetching technique that enhances the coverage of existing prefetchers. First, ADP consists of two prefetcher blocks. One is a path confidence-based lookahead prefetcher from prior work. The other is an LSTM prefetcher that predicts the next data address to be used based on block offset and page pattern history. The LSTM prefetcher is trained with the memory access patterns that the first prefetcher cannot predict, and thus it can further improve the performance of prefetching. Second, ADP adaptively selects a prefetcher which has a higher probability to hit the address of incoming memory accesses. Together, these features improve the coverage and accuracy of prefetching. In our analysis, we find that ADP achieves an IPC improvement of 24.2% over the confidence-based lookahead prefetcher.
- Open Access
LOSTIN: Logic Optimization via Spatio-Temporal Information with Hybrid Graph Models
Despite the stride made by machine learning (ML) based performance modeling, two major concerns that may impede production-ready ML applications in EDA are stringent accuracy requirements and generalization capability. To this end, we propose hybrid graph neural network (GNN) based approaches towards highly accurate quality-of-result (QoR) estimations with great generalization capability, specifically targeting logic synthesis optimization. The key idea is to simultaneously leverage spatio-temporal information from hardware designs and logic synthesis flows to forecast performance (i.e., delay/area) of various synthesis flows on different designs. The structural characteristics inside hardware designs are distilled and represented by GNNs; the temporal knowledge (i.e., relative ordering of logic transformations) in synthesis flows can be imposed on hardware designs by combining a virtually added supernode or a sequence processing model with conventional GNN models. Evaluation on 3.3 million data points shows that the testing mean absolute percentage error (MAPE) on designs seen and unseen during training are no more than 1.2% and 3.1%, respectively, which are 7-15X lower than existing studies.
Open AccessLOSTIN: Logic Optimization via Spatio-Temporal Information with Hybrid Graph Models
Despite the stride made by machine learning (ML) based performance modeling, two major concerns that may impede production-ready ML applications in EDA are stringent accuracy requirements and generalization capability. To this end, we propose hybrid graph neural network (GNN) based approaches towards highly accurate quality-of-result (QoR) estimations with great generalization capability, specifically targeting logic synthesis optimization. The key idea is to simultaneously leverage spatio-temporal information from hardware designs and logic synthesis flows to forecast performance (i.e., delay/area) of various synthesis flows on different designs. The structural characteristics inside hardware designs are distilled and represented by GNNs; the temporal knowledge (i.e., relative ordering of logic transformations) in synthesis flows can be imposed on hardware designs by combining a virtually added supernode or a sequence processing model with conventional GNN models. Evaluation on 3.3 million data points shows that the testing mean absolute percentage error (MAPE) on designs seen and unseen during training are no more than 1.2% and 3.1%, respectively, which are 7-15X lower than existing studies.
- 31 August 2022
- Published by The Research Institute of Police & Crime, Dongguk University in Korean Journal of Criminal Justice
- Vol. 5 (2) , 61-83
Galaxy Image Editing Tool Analysis to Detect Manipulated Images
AI 기술이 발달함에 따라 객체 지우기와 그림자 삭제와 같이 전문 기술이 요구되던 사 진 편집 기술을 일반인도 쉽게 활용할 수 있게 되었다. 해당 기술은 사용자들이 사진 에서 원치 않는 부분을 자연스럽게 지우거나 수정하도록 도와주지만, 접근성이 높아짐 에 따라 가짜뉴스 생성, 특정인 명예훼손 및 모욕, 증거 인멸과 같은 범죄에 악용될 여지도 커지고 있다. 이에 본 논문에서는 JPEG 파일 구조와 이미지의 그래픽적 특징, 애플리케이션 아티팩트를 종합적으로 고려하여 이미지 변조에 대응하는 기법을 제시한 다. 특히, 최근 삼성 갤럭시에서 기본 기능으로 제공하고 있는 AI 지우개 기능을 중점 적으로 분석하여 이미지의 변조 여부와 변조 영역을 판단하는 조사 방안을 제시한다.
- 31 August 2022
- Published by The Research Institute of Police & Crime, Dongguk University in Korean Journal of Criminal Justice
- Vol. 5 (2) , 61-83
Galaxy Image Editing Tool Analysis to Detect Manipulated Images
AI 기술이 발달함에 따라 객체 지우기와 그림자 삭제와 같이 전문 기술이 요구되던 사 진 편집 기술을 일반인도 쉽게 활용할 수 있게 되었다. 해당 기술은 사용자들이 사진 에서 원치 않는 부분을 자연스럽게 지우거나 수정하도록 도와주지만, 접근성이 높아짐 에 따라 가짜뉴스 생성, 특정인 명예훼손 및 모욕, 증거 인멸과 같은 범죄에 악용될 여지도 커지고 있다. 이에 본 논문에서는 JPEG 파일 구조와 이미지의 그래픽적 특징, 애플리케이션 아티팩트를 종합적으로 고려하여 이미지 변조에 대응하는 기법을 제시한 다. 특히, 최근 삼성 갤럭시에서 기본 기능으로 제공하고 있는 AI 지우개 기능을 중점 적으로 분석하여 이미지의 변조 여부와 변조 영역을 판단하는 조사 방안을 제시한다.
- 3 December 2021
- Published by International Display Workshops General Incorporated Association in Proceedings of the International Display Workshops
- p. 820
Disruptive Infrared Image Sensors Enabled by Quantum Dots
Proceedings of the International Display Workshops Volume 28 (IDW '21),Search for presentations
- 3 December 2021
- Published by International Display Workshops General Incorporated Association in Proceedings of the International Display Workshops
- p. 820
Disruptive Infrared Image Sensors Enabled by Quantum Dots
Proceedings of the International Display Workshops Volume 28 (IDW '21),Search for presentations