Inho Lee
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Inho Lee has a total of 18 co-authors. Here are the top co-authors by number of shared publications:
- Yongjun Park
Yonsei University, Seoul, Republic of Korea
4 shared publications - Beomseok Kim
Department of Intelligence and Information, Research Institute for Convergence Science, Inter-University Semiconductor Research Center, Seoul National University, Seoul, South Korea
1 shared publication - Bongjoon Hyun
KAIST, Daejeon, Republic of Korea
1 shared publication - Donghyeon Kim
Hanyang University, Seoul, Republic of Korea
1 shared publication - Dongsuk Jeon
Seoul National University
1 shared publication Hongjun Um
Department of Computer Science, Hanyang University, Seoul, South Korea
1 shared publication
Latest Publications
- Open Access
- 15 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 80-84
FLUX: Frequency Scaling with Layer-wise Utilization for Energy-Efficient NPU Execution (WIP)
With the widespread adoption of Deep Neural Networks (DNNs), Neural Processing Units (NPUs) are emerging as energy-efficient alternatives to GPUs through parallel processing and high data reuse. However, since diverse deep learning kernels have different memory and computation resource requirements, a utilization imbalance between memory and computation resources often occurs. To address this challenge, we propose FLUX, a frequency-scalable NPU system that applies Dynamic Frequency Scaling (DFS) to the core's frequency based on layer-wise arithmetic intensity. FLUX splits the clocking into an adjustable core domain and a fixed system domain. The core-domain frequency is first estimated via roofline cycle analysis and then refined at runtime by an Energy-Delay Product (EDP)-driven calibration. Evaluation on a Gemmini NPU using a 28nm process technology shows that FLUX achieves 5.7%, 16.5%, and 27.9% EDP improvements on 8×8, 16×16, and 32×32 systolic arrays for ResNet50, respectively, reaching within 1.7% of the oracle optimal frequency assignment.
- 15 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 80-84
Open AccessFLUX: Frequency Scaling with Layer-wise Utilization for Energy-Efficient NPU Execution (WIP)
With the widespread adoption of Deep Neural Networks (DNNs), Neural Processing Units (NPUs) are emerging as energy-efficient alternatives to GPUs through parallel processing and high data reuse. However, since diverse deep learning kernels have different memory and computation resource requirements, a utilization imbalance between memory and computation resources often occurs. To address this challenge, we propose FLUX, a frequency-scalable NPU system that applies Dynamic Frequency Scaling (DFS) to the core's frequency based on layer-wise arithmetic intensity. FLUX splits the clocking into an adjustable core domain and a fixed system domain. The core-domain frequency is first estimated via roofline cycle analysis and then refined at runtime by an Energy-Delay Product (EDP)-driven calibration. Evaluation on a Gemmini NPU using a 28nm process technology shows that FLUX achieves 5.7%, 16.5%, and 27.9% EDP improvements on 8×8, 16×16, and 32×32 systolic arrays for ResNet50, respectively, reaching within 1.7% of the oracle optimal frequency assignment.
- Open Access
- 17 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 63-77
PIM-CCA: An Efficient PIM Architecture with Optimized Integration of Configurable Functional Units
- 17 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 63-77
Open AccessPIM-CCA: An Efficient PIM Architecture with Optimized Integration of Configurable Functional Units
- Open Access
- 24 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 134143-134159
Efficient Image Super-Resolution Using Dynamic Quality Control With Recursive Model Structures
Nowadays, as the demand for accurate object detection (OD) applications is increasing, several attempts have been made to introduce convolutional neural network (CNN)-based super-resolution (SR) into these applications to further improve their target accuracy. OD systems require real-time processing because they are widely used in latency-critical applications such as autonomous driving, augmented reality, and surveillance cameras. However, due to their high computational and memory requirements, the introduction of SR in OD systems often makes real-time processing difficult. To reduce computation, one possible solution is to make the SR network size smaller, but this may not be the best solution because it lowers the quality of the reconstructed image. In addition, performing SR on the entire frame is inefficient because there are unnecessary background elements other than the objects that need to be detected within a single frame. Performing SR on specific regions of interest (ROIs) instead can improve the overall efficiency. Therefore, we propose an efficient dynamic quality control SR (DQC-SR) system that can dynamically adjust the inference rate according to the number of ROIs by introducing recursion and early-exit architectures into the baseline SR network. The DQC-SR network (DQC-SRNet) adopts a recursive architecture in which layers share identical parameters and an early-exit architecture that adjusts the network’s depth through bypassing paths. The recursive architecture can reduce memory consumption, while the early-exit architecture can adapt computation dynamically. With these two architectures, the DQCSR system enables real-time SR inference on edge devices. For evaluation, we conducted example scenarios based on license plate recognition (LPR) situations and demonstrated the effectiveness of the DQC-SR system.
- 24 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 134143-134159
Open AccessEfficient Image Super-Resolution Using Dynamic Quality Control With Recursive Model Structures
Nowadays, as the demand for accurate object detection (OD) applications is increasing, several attempts have been made to introduce convolutional neural network (CNN)-based super-resolution (SR) into these applications to further improve their target accuracy. OD systems require real-time processing because they are widely used in latency-critical applications such as autonomous driving, augmented reality, and surveillance cameras. However, due to their high computational and memory requirements, the introduction of SR in OD systems often makes real-time processing difficult. To reduce computation, one possible solution is to make the SR network size smaller, but this may not be the best solution because it lowers the quality of the reconstructed image. In addition, performing SR on the entire frame is inefficient because there are unnecessary background elements other than the objects that need to be detected within a single frame. Performing SR on specific regions of interest (ROIs) instead can improve the overall efficiency. Therefore, we propose an efficient dynamic quality control SR (DQC-SR) system that can dynamically adjust the inference rate according to the number of ROIs by introducing recursion and early-exit architectures into the baseline SR network. The DQC-SR network (DQC-SRNet) adopts a recursive architecture in which layers share identical parameters and an early-exit architecture that adjusts the network’s depth through bypassing paths. The recursive architecture can reduce memory consumption, while the early-exit architecture can adapt computation dynamically. With these two architectures, the DQCSR system enables real-time SR inference on edge devices. For evaluation, we conducted example scenarios based on license plate recognition (LPR) situations and demonstrated the effectiveness of the DQC-SR system.
- 17 June 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 30 (10) , 1461-1472
Dynamic Rate Neural Acceleration Using Multiprocessing Mode Support
Multiobject detection has become an integral component in various neural applications, such as autonomous driving and augmented reality. The system should be able to recognize and process multiple objects simultaneously. Moreover, the performance requirements for this system can be dynamically changed depending on the number of regions of interest (ROIs) in each frame. Consequently, the processing unit (PU) of the neural acceleration system should provide various inference rates. Therefore, we present a field-programmable gate array (FPGA)-based dynamic rate neural acceleration system called MultiLockOn to dynamically change the inference performance according to the number of ROIs per frame. It supports multiprocessing modes with different speeds through the introduction of novel multi-mode processing engines (PEs) comprising minimum reconfigurable interconnections across inference modes to minimize hardware overhead. The MultiLockOn system can provide an improvement of up to$4\times $ in the inference performance compared to that of DNNWeaver and$5.7\times $ compared to that of the ARM Cortex-A53 with minimum accuracy loss by supporting the multiprocessing modes.- 17 June 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 30 (10) , 1461-1472
Dynamic Rate Neural Acceleration Using Multiprocessing Mode Support
Multiobject detection has become an integral component in various neural applications, such as autonomous driving and augmented reality. The system should be able to recognize and process multiple objects simultaneously. Moreover, the performance requirements for this system can be dynamically changed depending on the number of regions of interest (ROIs) in each frame. Consequently, the processing unit (PU) of the neural acceleration system should provide various inference rates. Therefore, we present a field-programmable gate array (FPGA)-based dynamic rate neural acceleration system called MultiLockOn to dynamically change the inference performance according to the number of ROIs per frame. It supports multiprocessing modes with different speeds through the introduction of novel multi-mode processing engines (PEs) comprising minimum reconfigurable interconnections across inference modes to minimize hardware overhead. The MultiLockOn system can provide an improvement of up to$4\times $ in the inference performance compared to that of DNNWeaver and$5.7\times $ compared to that of the ARM Cortex-A53 with minimum accuracy loss by supporting the multiprocessing modes.