Tae Hee Han
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Tae Hee Han has a total of 67 co-authors. Here are the top co-authors by number of shared publications:
Chang Ho Ryu
5 shared publicationsYong Wook Kim
Department of Electrical and Computer Engineering, Sungkyunkwan University, Suwon, South Korea
5 shared publicationsJae Hoon Lee
Memory Division, Samsung Electronics, Hwaseong, South Korea
3 shared publications- Ju Yeon Kang
Department of Electrical and Computer Engineering, Sungkyunkwan University, Suwon, South Korea
3 shared publications Young Sik Lee
Department of Electrical and Computer Engineering, Sungkyunkwan University, Suwon, Republic of Korea
3 shared publications- Hyun Jo Lim
Department of Artificial Intelligence, Sungkyunkwan University, Suwon 16419, Republic of Korea
2 shared publications
Latest Publications
- 17 October 2026
- Published by Elsevier in Journal of Systems Architecture
- Vol. 181
Anchor-NDP: An importance-based prefix KV management framework for CXL-enabled NDP
- 17 October 2026
- Published by Elsevier in Journal of Systems Architecture
- Vol. 181
Anchor-NDP: An importance-based prefix KV management framework for CXL-enabled NDP
- Open Access
- 17 October 2026
- Published by Elsevier in Journal of Systems Architecture
- Vol. 181
SNAP: Spiking neural network-based adaptive data prefetcher via online learning
- 17 October 2026
- Published by Elsevier in Journal of Systems Architecture
- Vol. 181
Open AccessSNAP: Spiking neural network-based adaptive data prefetcher via online learning
- Open Access
- 27 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 1030-1040
Bypass-Enabled Topology-Agnostic Multicast Router for Fault-Tolerant Network-on-Chips
While dynamic rerouting represents the predominant fault-tolerant strategy for Network-on-Chip (NoC) multicast communications, circumventing permanent hardware failures inevitably concentrates detour traffic at specific intermediate nodes. Rerouting exacerbates pre-existing traffic skew and accelerates hardware aging — a problem that existing solutions fundamentally fail to address. Furthermore, as modern application-specific system-on-chips (SoCs) increasingly adopt irregular topologies, an effective fault-tolerant architecture must extend to such custom topologies. To address these challenges, we introduce BEAM, which integrates three key components: (1) A sequential topology embedding algorithm that employs a heuristic search with Warnsdorff’s rule to construct a near-Hamiltonian path, abstracting any arbitrary physical network into a unified logical sequence. (2) A neighbor-aware route distribution scheme that partitions multicast destinations into independent subsets based on router indices, enabling concurrent packet forwarding to improve throughput and fairness. (3) A lightweight bypass router architecture featuring a dedicated fault control unit that instantly redirects incoming packets to a physical bypass channel upon fault detection, circumventing faulty components without the overhead of full structural redundancy. Evaluation results demonstrate that under two fault conditions on different 64-node topologies, BEAM achieves 20.4% higher saturation throughput and 16.0% higher fairness on average than the evaluated schemes, with negligible hardware overhead.
- 27 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 1030-1040
Open AccessBypass-Enabled Topology-Agnostic Multicast Router for Fault-Tolerant Network-on-Chips
While dynamic rerouting represents the predominant fault-tolerant strategy for Network-on-Chip (NoC) multicast communications, circumventing permanent hardware failures inevitably concentrates detour traffic at specific intermediate nodes. Rerouting exacerbates pre-existing traffic skew and accelerates hardware aging — a problem that existing solutions fundamentally fail to address. Furthermore, as modern application-specific system-on-chips (SoCs) increasingly adopt irregular topologies, an effective fault-tolerant architecture must extend to such custom topologies. To address these challenges, we introduce BEAM, which integrates three key components: (1) A sequential topology embedding algorithm that employs a heuristic search with Warnsdorff’s rule to construct a near-Hamiltonian path, abstracting any arbitrary physical network into a unified logical sequence. (2) A neighbor-aware route distribution scheme that partitions multicast destinations into independent subsets based on router indices, enabling concurrent packet forwarding to improve throughput and fairness. (3) A lightweight bypass router architecture featuring a dedicated fault control unit that instantly redirects incoming packets to a physical bypass channel upon fault detection, circumventing faulty components without the overhead of full structural redundancy. Evaluation results demonstrate that under two fault conditions on different 64-node topologies, BEAM achieves 20.4% higher saturation throughput and 16.0% higher fairness on average than the evaluated schemes, with negligible hardware overhead.
- Open Access
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Phantom Walk: Reducing the Post-Migration Translation Path via Proactive Translation Dissemination in Multi-GPU Systems
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessPhantom Walk: Reducing the Post-Migration Translation Path via Proactive Translation Dissemination in Multi-GPU Systems
- Open Access
- 10 October 2025
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 81 (15) , 1-40
LNBN: layer-flexible non-blocking bypass network-on-chip for accelerating DNN inference
- 10 October 2025
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 81 (15) , 1-40
Open AccessLNBN: layer-flexible non-blocking bypass network-on-chip for accelerating DNN inference
- Open Access
- 2 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 115048-115062
BTCP: Binary Temporal Convolutional Network-Based Data Prefetcher for Low Inference Latency and Storage Overhead
Data prefetching is crucial for hiding long-latency memory access on modern high-performance processors. Machine learning (ML)-based data prefetchers have demonstrated superior address prediction performance compared with traditional rule-based prefetchers by learning complex patterns to hide long-latency memory access. However, prioritizing only predictive performance introduces challenges, such as increased inference latency and significant storage demands, when deployed in real hardware systems. To address these issues, we propose a binary temporal convolutional network-based data prefetcher (BTCP) that offers advantages in terms of computational efficiency and memory requirements, enabling feasible hardware implementation. The efficiency of BTCP is fundamentally attributed to its binary temporal convolutional network architecture, which achieves low inference latency and storage overhead by processing memory patterns through efficient binary operations. BTCP aids in detecting memory patterns by processing addresses and program counters through bitwise XOR operations and feeding them into a neural network. We employed positional delta maps to predict the closest future address, integrating address deltas with temporal weights for labeling. BTCP facilitates variable-degree prefetching by incorporating the ratio of irregular addresses into the positional delta map for enhanced training. The design choices result in an inference latency and storage overhead of only 134 cycles and 4,654 bytes, respectively, indicating a substantial decrease of 95.87% and 99.70%, respectively, compared to the leading ML-based prefetcher, TransFetch. With benchmark evaluations using SPEC CPU and GAP, BTCP achieved an IPC improvement of 18.70% over systems without prefetching, surpassing the state-of-the-art rule-based prefetcher Bingo by 2.78% and ML-based prefetcher TransFetch by 0.16%.
- 2 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 115048-115062
Open AccessBTCP: Binary Temporal Convolutional Network-Based Data Prefetcher for Low Inference Latency and Storage Overhead
Data prefetching is crucial for hiding long-latency memory access on modern high-performance processors. Machine learning (ML)-based data prefetchers have demonstrated superior address prediction performance compared with traditional rule-based prefetchers by learning complex patterns to hide long-latency memory access. However, prioritizing only predictive performance introduces challenges, such as increased inference latency and significant storage demands, when deployed in real hardware systems. To address these issues, we propose a binary temporal convolutional network-based data prefetcher (BTCP) that offers advantages in terms of computational efficiency and memory requirements, enabling feasible hardware implementation. The efficiency of BTCP is fundamentally attributed to its binary temporal convolutional network architecture, which achieves low inference latency and storage overhead by processing memory patterns through efficient binary operations. BTCP aids in detecting memory patterns by processing addresses and program counters through bitwise XOR operations and feeding them into a neural network. We employed positional delta maps to predict the closest future address, integrating address deltas with temporal weights for labeling. BTCP facilitates variable-degree prefetching by incorporating the ratio of irregular addresses into the positional delta map for enhanced training. The design choices result in an inference latency and storage overhead of only 134 cycles and 4,654 bytes, respectively, indicating a substantial decrease of 95.87% and 99.70%, respectively, compared to the leading ML-based prefetcher, TransFetch. With benchmark evaluations using SPEC CPU and GAP, BTCP achieved an IPC improvement of 18.70% over systems without prefetching, surpassing the state-of-the-art rule-based prefetcher Bingo by 2.78% and ML-based prefetcher TransFetch by 0.16%.
- Open Access
- 26 July 2024
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 19 (7) , e0306627
Deep learning-based material decomposition of iodine and calcium in mobile photon counting detector CT
Photon-counting detector (PCD)-based computed tomography (CT) offers several advantages over conventional energy-integrating detector-based CT. Among them, the ability to discriminate energy exhibits significant potential for clinical applications because it provides material-specific information. That is, material decomposition (MD) can be achieved through energy discrimination. In this study, deep learning-based material decomposition was performed using live animal data. We propose MD-Unet, which is a deep learning strategy for material decomposition based on an Unet architecture trained with data from three energy bins. To mitigate the data insufficiency, we developed a pretrained model incorporating various simulation data forms and augmentation strategies. Incorporating these approaches into model training results in enhanced precision in material decomposition, thereby enabling the identification of distinct materials at individual pixel locations. The trained network was applied to the acquired animal data to evaluate material decomposition results. Compared with conventional methods, the newly generated MD-Unet demonstrated more accurate material decomposition imaging. Moreover, the network demonstrated an improved material decomposition ability and significantly reduced noise. In addition, they can potentially offer an enhancement level similar to that of a typical contrast agent. This implies that it can acquire images of the same quality with fewer contrast agents administered to patients, thereby demonstrating its significant clinical value.
- 26 July 2024
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 19 (7) , e0306627
Open AccessDeep learning-based material decomposition of iodine and calcium in mobile photon counting detector CT
Photon-counting detector (PCD)-based computed tomography (CT) offers several advantages over conventional energy-integrating detector-based CT. Among them, the ability to discriminate energy exhibits significant potential for clinical applications because it provides material-specific information. That is, material decomposition (MD) can be achieved through energy discrimination. In this study, deep learning-based material decomposition was performed using live animal data. We propose MD-Unet, which is a deep learning strategy for material decomposition based on an Unet architecture trained with data from three energy bins. To mitigate the data insufficiency, we developed a pretrained model incorporating various simulation data forms and augmentation strategies. Incorporating these approaches into model training results in enhanced precision in material decomposition, thereby enabling the identification of distinct materials at individual pixel locations. The trained network was applied to the acquired animal data to evaluate material decomposition results. Compared with conventional methods, the newly generated MD-Unet demonstrated more accurate material decomposition imaging. Moreover, the network demonstrated an improved material decomposition ability and significantly reduced noise. In addition, they can potentially offer an enhancement level similar to that of a typical contrast agent. This implies that it can acquire images of the same quality with fewer contrast agents administered to patients, thereby demonstrating its significant clinical value.
- Open Access
- 4 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 49508-49527
OWBM: OSNR-Aware Wavelength Allocation and Branching Methods for Multicast Routing in Custom Topology-Based Optical Network-on-Chips
In light of the rapid advancements in big data and artificial intelligence applications, heterogeneous computing (HGC) platforms that integrate diverse computing units have gained traction with the aim of achieving energy efficiency and high performance. A custom topology-based optical network-on-chip (ONoC) that provides unparalleled diversity between computing nodes is expected to be the next-generation communication infrastructure for meeting the bandwidth and energy efficiency requirements of HGC. One of the recent challenges in the field of ONoCs is to accelerate multicast routing via wavelength division multiplexing (WDM), which dispatches data parallelly across non-interfering wavelengths. The optimization of network throughput and laser power efficiency revolves around two factors: the number of wavelengths and optical signal-to-noise ratio (OSNR). Accordingly, we introduce OSNR-aware wavelength allocation and branching methods for multicast routing (OWBM) tailored to an HGC platform in a customized ONoC. OWBM increases the wavelength resource efficiency by establishing independent routing paths in the partitioned destination nodes such that each routing path is guaranteed to be prevented from overlapping among the partitions. Moreover, the adaptive branching mechanism of OWBM adaptively selects path-based routing and OSNR-aware routing on the fly according to the wavelength allocation cases, further augmenting the throughput and laser power efficiency. Consequently, OWBM outperformed conventional tree- and path-based multicast approaches by elevating the average throughput by 47.39% and curbing the laser power consumption by up to 35.92% in various convolutional neural network benchmarks. Compared with the existing ONoC wavelength allocation techniques, the OWBM demonstrated a maximum of 42.46% enhanced wavelength utilization on a 64-core HGC platform.
- 4 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 49508-49527
Open AccessOWBM: OSNR-Aware Wavelength Allocation and Branching Methods for Multicast Routing in Custom Topology-Based Optical Network-on-Chips
In light of the rapid advancements in big data and artificial intelligence applications, heterogeneous computing (HGC) platforms that integrate diverse computing units have gained traction with the aim of achieving energy efficiency and high performance. A custom topology-based optical network-on-chip (ONoC) that provides unparalleled diversity between computing nodes is expected to be the next-generation communication infrastructure for meeting the bandwidth and energy efficiency requirements of HGC. One of the recent challenges in the field of ONoCs is to accelerate multicast routing via wavelength division multiplexing (WDM), which dispatches data parallelly across non-interfering wavelengths. The optimization of network throughput and laser power efficiency revolves around two factors: the number of wavelengths and optical signal-to-noise ratio (OSNR). Accordingly, we introduce OSNR-aware wavelength allocation and branching methods for multicast routing (OWBM) tailored to an HGC platform in a customized ONoC. OWBM increases the wavelength resource efficiency by establishing independent routing paths in the partitioned destination nodes such that each routing path is guaranteed to be prevented from overlapping among the partitions. Moreover, the adaptive branching mechanism of OWBM adaptively selects path-based routing and OSNR-aware routing on the fly according to the wavelength allocation cases, further augmenting the throughput and laser power efficiency. Consequently, OWBM outperformed conventional tree- and path-based multicast approaches by elevating the average throughput by 47.39% and curbing the laser power consumption by up to 35.92% in various convolutional neural network benchmarks. Compared with the existing ONoC wavelength allocation techniques, the OWBM demonstrated a maximum of 42.46% enhanced wavelength utilization on a 64-core HGC platform.
- Open Access
- 25 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 145238-145253
RL-Based Cache Replacement: A Modern Interpretation of Belady’s Algorithm With Bypass Mechanism and Access Type Analysis
Belady’s algorithm is widely known as an optimal cache replacement policy. It has been the foundation of numerous recent studies on cache replacement policies, and most studies assume this as an upper limit. Despite its widespread adoption, we discovered opportunities to unleash the headroom by addressing cache access types and implementing cache bypass. In this study, we propose Stormbird, a cache replacement policy that synergistically integrates the extensions of Belady’s algorithm and the power of reinforcement learning. Reinforcement learning is well-suited for cache replacement policy problems owing to its ability to interact dynamically with the environment, adapt to changing access patterns, and optimize the maximum cumulative rewards. Stormbird utilizes several selected features from the reinforcement learning model to enhance the instructions per cycle efficiency while maintaining a low hardware overhead. Furthermore, it considers cache access types and integrates dynamic set dueling techniques to improve the cache performance. For 2 MB last-level cache per core, Stormbird achieves an average instructions per cycle improvement of 0.13% over the previous state-of-the-art on a single-core system and 0.02% on a four-core system while simultaneously reducing hardware overhead by 62.5%. Stormbird incurs a low hardware overhead of only 10.5 KB for 2 MB last-level cache and can be implemented without using program counter values.
- 25 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 145238-145253
Open AccessRL-Based Cache Replacement: A Modern Interpretation of Belady’s Algorithm With Bypass Mechanism and Access Type Analysis
Belady’s algorithm is widely known as an optimal cache replacement policy. It has been the foundation of numerous recent studies on cache replacement policies, and most studies assume this as an upper limit. Despite its widespread adoption, we discovered opportunities to unleash the headroom by addressing cache access types and implementing cache bypass. In this study, we propose Stormbird, a cache replacement policy that synergistically integrates the extensions of Belady’s algorithm and the power of reinforcement learning. Reinforcement learning is well-suited for cache replacement policy problems owing to its ability to interact dynamically with the environment, adapt to changing access patterns, and optimize the maximum cumulative rewards. Stormbird utilizes several selected features from the reinforcement learning model to enhance the instructions per cycle efficiency while maintaining a low hardware overhead. Furthermore, it considers cache access types and integrates dynamic set dueling techniques to improve the cache performance. For 2 MB last-level cache per core, Stormbird achieves an average instructions per cycle improvement of 0.13% over the previous state-of-the-art on a single-core system and 0.02% on a four-core system while simultaneously reducing hardware overhead by 62.5%. Stormbird incurs a low hardware overhead of only 10.5 KB for 2 MB last-level cache and can be implemented without using program counter values.
- Open Access
- 13 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 128306-128315
SFAO: Sign-Flipping-Aware Optimization for Early-Stopping of Binarized Neural Networks
One of the vital challenges for the binary neural networks (BNNs) is improving their inference performance by expanding their data representation capabilities for figuring out delicate patterns and nuances in the data. Addressing the explosive computational demands on neural network training is essential to guarantee sustainable development and scalable deployment. However, mitigating the increase in the computational cost during the training phase is critical for ensuring sustainability and scalability during deployment. In this study, an advanced sign-flipping-aware optimizer (SFAO) that focuses on BNNs was introduced to diminish the computational burden. SFAO balanced the model performance and computational cost through sign-flipping-aware updating rules throughout the training of BNNs. SFAO optimizer, tailored for BNNs with binary weight-specific updating rules, considerably reduced the computing resources needed for training on the CIFAR-10 dataset. Specifically, it surpassed the conventional full-precision updating rule by reducing the total instruction count by 21.89%. In contrast, SFAO showed a marginal 0.44% decline in the image classification accuracy relative to the updating rules for the full-precision parameters. Furthermore, the implementation of early stopping using the sign flip rate led to a notable reduction of 9.37% in the average computation time per network for the ImageNet dataset.
- 13 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 128306-128315
Open AccessSFAO: Sign-Flipping-Aware Optimization for Early-Stopping of Binarized Neural Networks
One of the vital challenges for the binary neural networks (BNNs) is improving their inference performance by expanding their data representation capabilities for figuring out delicate patterns and nuances in the data. Addressing the explosive computational demands on neural network training is essential to guarantee sustainable development and scalable deployment. However, mitigating the increase in the computational cost during the training phase is critical for ensuring sustainability and scalability during deployment. In this study, an advanced sign-flipping-aware optimizer (SFAO) that focuses on BNNs was introduced to diminish the computational burden. SFAO balanced the model performance and computational cost through sign-flipping-aware updating rules throughout the training of BNNs. SFAO optimizer, tailored for BNNs with binary weight-specific updating rules, considerably reduced the computing resources needed for training on the CIFAR-10 dataset. Specifically, it surpassed the conventional full-precision updating rule by reducing the total instruction count by 21.89%. In contrast, SFAO showed a marginal 0.44% decline in the image classification accuracy relative to the updating rules for the full-precision parameters. Furthermore, the implementation of early stopping using the sign flip rate led to a notable reduction of 9.37% in the average computation time per network for the ImageNet dataset.