Eojin Lee
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Eojin Lee has a total of 42 co-authors. Here are the top co-authors by number of shared publications:
- Jung Ho Ahn
Seoul National University, Seoul, Republic of Korea
12 shared publications - Jaehyun Park
Seoul National University, Seoul, Republic of Korea
6 shared publications - Byeongho Kim
Seoul National University, Seoul, South Korea
5 shared publications - Hwayong Nam
Seoul National University
4 shared publications - Hyeonu Kim
Seoul National University, Seoul, Republic of Korea
4 shared publications - Sungmin Yun
Seoul National University, Seoul, South Korea
4 shared publications
Latest Publications
- Open Access
- 5 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 893-905
Aurora: A Disaggregated GPU-PNM-PIM System for High-Throughput Mixed-Length LLM Inference
- 5 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 893-905
Open AccessAurora: A Disaggregated GPU-PNM-PIM System for High-Throughput Mixed-Length LLM Inference
- Open Access
- 5 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 795-810
Clone: A Collaborative Multi-device System for Retrieval-Augmented Generation over CXL
- 5 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 795-810
Open AccessClone: A Collaborative Multi-device System for Retrieval-Augmented Generation over CXL
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 25 (2) , 359-362
DYCA-HBM: Accelerating Long-Context LLM on HBM-PIM With Dynamic Channel-Allocation
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 25 (2) , 359-362
DYCA-HBM: Accelerating Long-Context LLM on HBM-PIM With Dynamic Channel-Allocation
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 25 (2) , 263-266
SPIRE-HBM: A Sub-Pseudo-Channel Architecture for HBM4 Bandwidth Scaling
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 25 (2) , 263-266
SPIRE-HBM: A Sub-Pseudo-Channel Architecture for HBM4 Bandwidth Scaling
- Open Access
- 17 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 49-62
ComPASS: A Compatible PIM Protocol Architecture and Scheduling Solution for Processor-PIM Collaboration
- 17 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 49-62
Open AccessComPASS: A Compatible PIM Protocol Architecture and Scheduling Solution for Processor-PIM Collaboration
- 2 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 24 (1) , 185-188
Fold-PIM: A Cost-Efficient LPDDR5-Based PIM for On-Device SLMs
The increasing demand for on-device AI applications has shifted focus to Small Language Models (SLMs) optimized for mobile environments. However, the limited memory bandwidth of LPDDR5-based systems presents significant challenges for efficiently executing memory-bound matrix-vector multiplication operations, a core component of SLM inference. In this paper, we propose Fold-PIM, an LPDDR5-based Processing-in-Memory (PIM) architecture designed to address these challenges. FoldPIM features a shared PU architecture that leverages subarray level parallelism while maintaining cost efficiency. Additionally, we introduce key techniques, including adaptive tiling for efficient GEMV operations across diverse matrix sizes, a transposed data layout to minimize buffer overhead, and a tailored protocol to reduce vector replacement latency. Our evaluation results demonstrate that Fold-PIM achieves up to 3.9× speedup of token generation time in SLM inference compared to the baseline system without PIM.
- 2 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 24 (1) , 185-188
Fold-PIM: A Cost-Efficient LPDDR5-Based PIM for On-Device SLMs
The increasing demand for on-device AI applications has shifted focus to Small Language Models (SLMs) optimized for mobile environments. However, the limited memory bandwidth of LPDDR5-based systems presents significant challenges for efficiently executing memory-bound matrix-vector multiplication operations, a core component of SLM inference. In this paper, we propose Fold-PIM, an LPDDR5-based Processing-in-Memory (PIM) architecture designed to address these challenges. FoldPIM features a shared PU architecture that leverages subarray level parallelism while maintaining cost efficiency. Additionally, we introduce key techniques, including adaptive tiling for efficient GEMV operations across diverse matrix sizes, a transposed data layout to minimize buffer overhead, and a tailored protocol to reduce vector replacement latency. Our evaluation results demonstrate that Fold-PIM achieves up to 3.9× speedup of token generation time in SLM inference compared to the baseline system without PIM.
- Open Access
- 30 March 2025
- Published by Association for Computing Machinery (ACM)
- p. 854-869
PET: Proactive Demotion for Efficient Tiered Memory Management
- 30 March 2025
- Published by Association for Computing Machinery (ACM)
- p. 854-869
Open AccessPET: Proactive Demotion for Efficient Tiered Memory Management
- 24 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 23 (2) , 183-186
Architecting Compatible PIM Protocol for CPU-PIM Collaboration
Processing in Memory (PIM) technology is gaining traction with the introduction of several prototype products. However, the interfaces of existing PIM devices hinder CPU performance excessively by delaying normal memory requests for long periods during PIM operations. In this paper, we propose a new PIM command and protocol designed for compatibility across various PIM devices and host processors, focusing on DRAM standards with limited command space. Our proposed command, PIM-ACT, activates multiple banks simultaneously with assigning the specific PIM operation. It closely follows the functionality of the ACT command for straightforward control by the memory controller. We also explore memory scheduling policies that balance the latency of conventional memory requests with the throughput of PIM workloads. Our evaluation demonstrates the effectiveness of our approach in optimizing both PIM and conventional workload performance.
- 24 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 23 (2) , 183-186
Architecting Compatible PIM Protocol for CPU-PIM Collaboration
Processing in Memory (PIM) technology is gaining traction with the introduction of several prototype products. However, the interfaces of existing PIM devices hinder CPU performance excessively by delaying normal memory requests for long periods during PIM operations. In this paper, we propose a new PIM command and protocol designed for compatibility across various PIM devices and host processors, focusing on DRAM standards with limited command space. Our proposed command, PIM-ACT, activates multiple banks simultaneously with assigning the specific PIM operation. It closely follows the functionality of the ACT command for straightforward control by the memory controller. We also explore memory scheduling policies that balance the latency of conventional memory requests with the throughput of PIM workloads. Our evaluation demonstrates the effectiveness of our approach in optimizing both PIM and conventional workload performance.
- Open Access
- 3 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 69-82
IDT: Intelligent Data Placement for Multi-tiered Main Memory with Reinforcement Learning
To address the limitation of a DRAM-based single-tier in satisfying the comprehensive demands of main memory, multi-tiered memory systems are gaining widespread adoption. To support these systems, operating-system-level solutions that analyze the application's memory access patterns and ensure data placement in the appropriate memory tier have been vastly explored. In this paper, we identify reinforcement learning (RL) as an effective solution for tiered memory management, and its policy can be formulated in a solvable form using RL. We also demonstrate that an effective region-granularity memory access monitoring method is necessary to provide an accurate environment state to the RL model. Thus, we propose IDT, an intelligent data placement for multi-tiered main memory. IDT incorporates an RL-based demotion policy autotuning and a mechanism that efficiently demotes cold pages to lower-tier memory. IDT also promotes hot pages to upper-tier memory to minimize access on slow memory, featuring a lightweight machine learning algorithm. IDT employs region-granularity memory access monitoring with statistical-testing-based adjacent region merge and split to improve precision and mitigate ambiguity observed in prior works. Experiments on an actual four-tiered memory system show that IDT achieves an average 2.08× speedup over the default Linux kernel and 11.2% performance improvement compared to the state-of-the-art solution.
- 3 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 69-82
Open AccessIDT: Intelligent Data Placement for Multi-tiered Main Memory with Reinforcement Learning
To address the limitation of a DRAM-based single-tier in satisfying the comprehensive demands of main memory, multi-tiered memory systems are gaining widespread adoption. To support these systems, operating-system-level solutions that analyze the application's memory access patterns and ensure data placement in the appropriate memory tier have been vastly explored. In this paper, we identify reinforcement learning (RL) as an effective solution for tiered memory management, and its policy can be formulated in a solvable form using RL. We also demonstrate that an effective region-granularity memory access monitoring method is necessary to provide an accurate environment state to the RL model. Thus, we propose IDT, an intelligent data placement for multi-tiered main memory. IDT incorporates an RL-based demotion policy autotuning and a mechanism that efficiently demotes cold pages to lower-tier memory. IDT also promotes hot pages to upper-tier memory to minimize access on slow memory, featuring a lightweight machine learning algorithm. IDT employs region-granularity memory access monitoring with statistical-testing-based adjacent region merge and split to improve precision and mitigate ambiguity observed in prior works. Experiments on an actual four-tiered memory system show that IDT achieves an average 2.08× speedup over the default Linux kernel and 11.2% performance improvement compared to the state-of-the-art solution.
- Open Access
- 30 May 2024
- Published by Association for Computing Machinery (ACM)
- p. 338-351
CLAY: CXL-based Scalable NDP Architecture Accelerating Embedding Layers
An embedding layer is one of the most critical building blocks of deep neural networks, especially for recommender systems and graph neural networks. The embedding layer dominates a large portion of the total execution time due to its large memory requirements and little data reuse in operations. To accelerate the embedding layers, dual in-line memory module (DIMM) based near-data processing architectures have been proposed. They amplify bandwidth by adding a processing unit to the DIMM’s buffer. However, prior architectures have less capacity scalability due to the limited number of memory channels. Crucially, they are limited in performance improvement due to the load imbalance problem and the limitations of DIMM-based memory systems with a multi-drop bus structure between the processing units and the host. In this paper, we propose CLAY, a CXL-based scalable near-data processing architecture that accelerates general embedding layers in DNN. Breaking away from conventional memory channel structures, CLAY interconnects the DRAM modules to reduce the data transfer overhead among DRAM modules. Furthermore, we devise a dedicated memory address mapping to mitigate load imbalance in CLAY and a packet duplication scheme that enables full utilization of CLAY by reducing the required instruction transmission bandwidth. We propose a method of scaling CLAY and a software stack to use CLAY. Compared to the state-of-the-art NDP architectures of FeaNMP and G-NMP, CLAY achieves an end-to-end speedup of up to 1.87 × and 2.77 × for recommender systems and graph neural networks, respectively.
- 30 May 2024
- Published by Association for Computing Machinery (ACM)
- p. 338-351
Open AccessCLAY: CXL-based Scalable NDP Architecture Accelerating Embedding Layers
An embedding layer is one of the most critical building blocks of deep neural networks, especially for recommender systems and graph neural networks. The embedding layer dominates a large portion of the total execution time due to its large memory requirements and little data reuse in operations. To accelerate the embedding layers, dual in-line memory module (DIMM) based near-data processing architectures have been proposed. They amplify bandwidth by adding a processing unit to the DIMM’s buffer. However, prior architectures have less capacity scalability due to the limited number of memory channels. Crucially, they are limited in performance improvement due to the load imbalance problem and the limitations of DIMM-based memory systems with a multi-drop bus structure between the processing units and the host. In this paper, we propose CLAY, a CXL-based scalable near-data processing architecture that accelerates general embedding layers in DNN. Breaking away from conventional memory channel structures, CLAY interconnects the DRAM modules to reduce the data transfer overhead among DRAM modules. Furthermore, we devise a dedicated memory address mapping to mitigate load imbalance in CLAY and a packet duplication scheme that enables full utilization of CLAY by reducing the required instruction transmission bandwidth. We propose a method of scaling CLAY and a software stack to use CLAY. Compared to the state-of-the-art NDP architectures of FeaNMP and G-NMP, CLAY achieves an end-to-end speedup of up to 1.87 × and 2.77 × for recommender systems and graph neural networks, respectively.