Jueun Jung
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Jueun Jung has a total of 17 co-authors. Here are the top co-authors by number of shared publications:
- Kyuho Jason Lee
Yonsei University
15 shared publications - Seungbin Kim
Ulsan National Institute of Science and Technology
9 shared publications - Bokyoung Seo
Ulsan National Institute of Science and Technology
8 shared publications - Donghyeon Han
Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology (MIT), Cambridge, MA, USA
4 shared publications - Ghangmin Yun
Yonsei University
4 shared publications - Junghyun Yoo
Yonsei University
3 shared publications
Latest Publications
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 73 (11) , 8015-8028
A 135-FPS and 0.81-mJ/frame Heterogeneous Processor for Real-Time and Energy-Efficient LiDAR Odometry and Mapping
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 73 (11) , 8015-8028
A 135-FPS and 0.81-mJ/frame Heterogeneous Processor for Real-Time and Energy-Efficient LiDAR Odometry and Mapping
- Open Access
- 5 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
HotBa: A Heterogeneous Mamba Accelerator with Δ-Guided Early Rejection for Speculative Decoding
Mamba has emerged as a promising alternative to Transformers for on-device LLM inference, offering linear complexity and eliminating KV-cache. However, autoregressive decoding reloads full model weights every token, accounting for 93.2% of total inference energy, with no existing accelerator resolving this overhead. We present HotBa, a heterogeneous Mamba accelerator that reduces per-token weight transfer by 82% and redundant computation by 59% via Δ-guided early rejection for wide-tree speculative decoding in Mamba, while a heterogeneous INT8/FP16 core and tree management unit achieve 40.4× area efficiency and 5.18× SSM speedup with 0.4% area overhead. Synthesized in 28nm CMOS, HotBa achieves 75.32 tokens/s with 1.65× speedup and 7.55× energy efficiency over the state-of-the-art Mamba accelerator.
- 5 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessHotBa: A Heterogeneous Mamba Accelerator with Δ-Guided Early Rejection for Speculative Decoding
Mamba has emerged as a promising alternative to Transformers for on-device LLM inference, offering linear complexity and eliminating KV-cache. However, autoregressive decoding reloads full model weights every token, accounting for 93.2% of total inference energy, with no existing accelerator resolving this overhead. We present HotBa, a heterogeneous Mamba accelerator that reduces per-token weight transfer by 82% and redundant computation by 59% via Δ-guided early rejection for wide-tree speculative decoding in Mamba, while a heterogeneous INT8/FP16 core and tree management unit achieve 40.4× area efficiency and 5.18× SSM speedup with 0.4% area overhead. Synthesized in 28nm CMOS, HotBa achieves 75.32 tokens/s with 1.65× speedup and 7.55× energy efficiency over the state-of-the-art Mamba accelerator.
- 1 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. PP (00189200) , 1-15
SERAH-CIM: Sparsity-Aware Effective Row Activation Analog-Digital Hybrid eDRAM CIM With In-Macro Multi-Row-Multi-Task Control
This article presents a sparsity-aware analog-digital hybrid embedded dynamic random access memory (eDRAM) computing-in-memory (CIM) processor for highly energy-efficient deep neural network (DNN) acceleration. Although CIM architectures execute multiplication-and-accumulation (MAC) more efficiently than von Neumann architectures, their practical energy efficiency for DNN acceleration remains limited due to several challenges. First, prior CIMs struggled to exploit massive sparsity because of their highly parallelized structures. Second, CIMs typically adopt computation in either the analog or the digital domain, facing fundamental trade-offs between analog-to-digital converter (ADC) overhead and throughput. Third, system throughput is degraded due to 1) workload imbalance among CIM macros during sparsity-aware computation caused by a random sparsity pattern, which ruins CIM macro utilization and 2)frequent refresh and weight-update operations that have plagued prior eDRAM CIMs. To address these challenges, the proposed eDRAM CIM processor introduces four key features: 1) input activation (IA) grouping convolution, which completely skips zero-weight computations by activating only the effective rows of the CIM macro, increasing the effective computation ratio by $4.59\times $ ; 2) a hybrid-CIM macro integrated with SAR-Flash ADC (SF-ADC) and reversed-MAC near-memory logic (RM-NML) for energy-efficient MAC operations in both the analog and the digital domains, improving macro efficiency by $2.39\times $ ; 3) sparsity-aware proactive scheduling (SPS) to maximize CIM macro utilization, reducing system latency by 10.4%; and 4)in-macro Multi-row-Multi-task (MRMT) control that enables concurrent refresh/update during in-memory computation, resulting in a 22.0% reduction in system latency and a $1.3\times $ increase in system energy efficiency. Fabricated in a 28 nm CMOS process, the proposed processor demonstrates high-energy efficiency across various benchmarks, outperforming previous CIM processors by $1.55\times $ and $10.37\times $ for ResNet-18 and VGGNet-16, respectively.
- 1 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. PP (00189200) , 1-15
SERAH-CIM: Sparsity-Aware Effective Row Activation Analog-Digital Hybrid eDRAM CIM With In-Macro Multi-Row-Multi-Task Control
This article presents a sparsity-aware analog-digital hybrid embedded dynamic random access memory (eDRAM) computing-in-memory (CIM) processor for highly energy-efficient deep neural network (DNN) acceleration. Although CIM architectures execute multiplication-and-accumulation (MAC) more efficiently than von Neumann architectures, their practical energy efficiency for DNN acceleration remains limited due to several challenges. First, prior CIMs struggled to exploit massive sparsity because of their highly parallelized structures. Second, CIMs typically adopt computation in either the analog or the digital domain, facing fundamental trade-offs between analog-to-digital converter (ADC) overhead and throughput. Third, system throughput is degraded due to 1) workload imbalance among CIM macros during sparsity-aware computation caused by a random sparsity pattern, which ruins CIM macro utilization and 2)frequent refresh and weight-update operations that have plagued prior eDRAM CIMs. To address these challenges, the proposed eDRAM CIM processor introduces four key features: 1) input activation (IA) grouping convolution, which completely skips zero-weight computations by activating only the effective rows of the CIM macro, increasing the effective computation ratio by $4.59\times $ ; 2) a hybrid-CIM macro integrated with SAR-Flash ADC (SF-ADC) and reversed-MAC near-memory logic (RM-NML) for energy-efficient MAC operations in both the analog and the digital domains, improving macro efficiency by $2.39\times $ ; 3) sparsity-aware proactive scheduling (SPS) to maximize CIM macro utilization, reducing system latency by 10.4%; and 4)in-macro Multi-row-Multi-task (MRMT) control that enables concurrent refresh/update during in-memory computation, resulting in a 22.0% reduction in system latency and a $1.3\times $ increase in system energy efficiency. Fabricated in a 28 nm CMOS process, the proposed processor demonstrates high-energy efficiency across various benchmarks, outperforming previous CIM processors by $1.55\times $ and $10.37\times $ for ResNet-18 and VGGNet-16, respectively.
- 7 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems II: Express Briefs
- Vol. 72 (9) , 1273-1277
An Energy-Efficient Image Deblurring Accelerator With Quad-Base-Quad-Scale Quantized Format and Layer Normalization-Aware Optimization
This brief proposes a novel data-format-based image deblurring accelerator with layer normalization and UNet architecture optimization for mobile cameras. As the demand for photography in dynamic environments continues to grow and the limitations of physical stabilization are tightening, post-processing methods to restore sharp images have gained increasing attention, notably deblurring methods based on convolutional neural networks. However, their heavy computational cost hinders their integration into mobile computing platforms. The proposed accelerator enables energy-efficient acceleration of deblurring through the following three key features: 1) A Quad-base-Quad-scale Quantized format that maintains image quality with only 8-bit, reducing external memory access (EMA) by 33% and achieving 75.7% higher multiply-and-accumulation (MAC) energy efficiency compared to conventional 12-bit precision; 2) A Layer Normalization-Aware Optimization technique, enabling parallel normalization and fusion of affine transformation; 3) A dual-stationary systolic array architecture that selects the optimal dataflow for each UNet block based on processing element (PE) utilization. As a result, the proposed accelerator achieves 2.49 TOPS/W, which is $2.23\times $ higher than prior work, enabling energy-efficient deblurring for mobile applications.
- 7 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems II: Express Briefs
- Vol. 72 (9) , 1273-1277
An Energy-Efficient Image Deblurring Accelerator With Quad-Base-Quad-Scale Quantized Format and Layer Normalization-Aware Optimization
This brief proposes a novel data-format-based image deblurring accelerator with layer normalization and UNet architecture optimization for mobile cameras. As the demand for photography in dynamic environments continues to grow and the limitations of physical stabilization are tightening, post-processing methods to restore sharp images have gained increasing attention, notably deblurring methods based on convolutional neural networks. However, their heavy computational cost hinders their integration into mobile computing platforms. The proposed accelerator enables energy-efficient acceleration of deblurring through the following three key features: 1) A Quad-base-Quad-scale Quantized format that maintains image quality with only 8-bit, reducing external memory access (EMA) by 33% and achieving 75.7% higher multiply-and-accumulation (MAC) energy efficiency compared to conventional 12-bit precision; 2) A Layer Normalization-Aware Optimization technique, enabling parallel normalization and fusion of affine transformation; 3) A dual-stationary systolic array architecture that selects the optimal dataflow for each UNet block based on processing element (PE) utilization. As a result, the proposed accelerator achieves 2.49 TOPS/W, which is $2.23\times $ higher than prior work, enabling energy-efficient deblurring for mobile applications.
- 12 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 60 (10) , 3655-3664
A LiDAR–PNN Pipelined Processor With Cylindrical Bin Partitioning and Halo Indexing for 3-D Perception in Outdoor Autonomous Driving Applications
With the rapid attention of real-time 3-D perception in outdoor autonomous driving applications, 3-D point-cloud neural network (PNN) using LiDAR data has been actively developed to capture the accurate semantic surrounding information. Previous 3-D PNN processors solely focused on the acceleration of algorithm structure [e.g., nearest neighbor searching, shared multi-layer perceptron (MLP), etc.] required due to the unordered and unstructured characteristics of point-cloud with the assumption that the entire point-cloud data are stored in external memory. However, they were impractical for real-time PNN in outdoor environments with LiDAR, because the LiDAR takes ~100 ms of sensing latency to obtain the 360° fully scanned LiDAR points, causing a waste of system latency and additional delays due to external memory accesses. Moreover, their Cartesian-based bin partitioning method exacerbates workload imbalance in outdoor applications due to spatial sparsity as the point distribution becomes increasingly sparse with distance from the sensor, causing severe performance degradation. To address these challenges, this paper proposes an efficient 3-D PNN processor with LiDAR–PNN processor pipelined structure that eliminates the external memory between LiDAR and processor and hides sensing latency behind processing time by utilizing the mechanical characteristic of LiDAR with the following hardware features: 1) cylindrical partitioning core with novel cylindrical bin partitioning method as well as halo indexing to process with the partially scanned ( $2{\theta }$ ) data and mitigate the PNN accuracy loss; 2) pseudo-random number generator-based sampling unit and k-nearest neighbor searching cores with unified neighbor searching algorithm to reduce the computation cost; and 3) linked-list memory management unit and predicted memory allocator to efficiently manage the point-cloud and support the parallel processing of neighbor searching and shared MLP. As a result, the proposed PNN processor achieves 2.18 M point/s of peak performance and $0.40~{\mu }$ J/point of energy with 108 K points.
- 12 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 60 (10) , 3655-3664
A LiDAR–PNN Pipelined Processor With Cylindrical Bin Partitioning and Halo Indexing for 3-D Perception in Outdoor Autonomous Driving Applications
With the rapid attention of real-time 3-D perception in outdoor autonomous driving applications, 3-D point-cloud neural network (PNN) using LiDAR data has been actively developed to capture the accurate semantic surrounding information. Previous 3-D PNN processors solely focused on the acceleration of algorithm structure [e.g., nearest neighbor searching, shared multi-layer perceptron (MLP), etc.] required due to the unordered and unstructured characteristics of point-cloud with the assumption that the entire point-cloud data are stored in external memory. However, they were impractical for real-time PNN in outdoor environments with LiDAR, because the LiDAR takes ~100 ms of sensing latency to obtain the 360° fully scanned LiDAR points, causing a waste of system latency and additional delays due to external memory accesses. Moreover, their Cartesian-based bin partitioning method exacerbates workload imbalance in outdoor applications due to spatial sparsity as the point distribution becomes increasingly sparse with distance from the sensor, causing severe performance degradation. To address these challenges, this paper proposes an efficient 3-D PNN processor with LiDAR–PNN processor pipelined structure that eliminates the external memory between LiDAR and processor and hides sensing latency behind processing time by utilizing the mechanical characteristic of LiDAR with the following hardware features: 1) cylindrical partitioning core with novel cylindrical bin partitioning method as well as halo indexing to process with the partially scanned ( $2{\theta }$ ) data and mitigate the PNN accuracy loss; 2) pseudo-random number generator-based sampling unit and k-nearest neighbor searching cores with unified neighbor searching algorithm to reduce the computation cost; and 3) linked-list memory management unit and predicted memory allocator to efficiently manage the point-cloud and support the parallel processing of neighbor searching and shared MLP. As a result, the proposed PNN processor achieves 2.18 M point/s of peak performance and $0.40~{\mu }$ J/point of energy with 108 K points.
- 21 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Micro
- Vol. 45 (2) , 67-77
A Mobile Semantic Lidar SLAM Processor With Artificial-Intelligence-Based 3-D Perception and Spatiotemporal-Aware Computing
A low-power artificial intelligence (AI)-based semantic LiDAR SLAM processor is proposed to expand autonomous driving into emerging mobile robots. It combines point neural network (PNN)-based 3D segmentation with LiDAR SLAM to minimize pose errors due to the lack of perception in previous SLAM. The proposed processor is designed with a heterogeneous multi-core architecture utilizing SIMD and reconfigurable processing elements to fully support the three main operations: k-nearest neighbor (kNN); PNN; and non-linear optimization. It accelerates kNN operations with spherical bin partitioning optimized for the distribution of LiDAR data to eliminate unnecessary search spaces. In addition, the proposed spatio-temporal-aware computing minimizes excessive memory overhead and workload imbalance in kNN and PNN operations. Consequently, fabricated with 28-nm CMOS technology, the processor achieves 8.245 mJ/frame of energy consumption and a maximum performance of 20.7 ms latency, successfully demonstrating real-time semantic LiDAR SLAM system with 99.86% lower power consumption compared to modern CPU+GPU platforms.
- 21 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Micro
- Vol. 45 (2) , 67-77
A Mobile Semantic Lidar SLAM Processor With Artificial-Intelligence-Based 3-D Perception and Spatiotemporal-Aware Computing
A low-power artificial intelligence (AI)-based semantic LiDAR SLAM processor is proposed to expand autonomous driving into emerging mobile robots. It combines point neural network (PNN)-based 3D segmentation with LiDAR SLAM to minimize pose errors due to the lack of perception in previous SLAM. The proposed processor is designed with a heterogeneous multi-core architecture utilizing SIMD and reconfigurable processing elements to fully support the three main operations: k-nearest neighbor (kNN); PNN; and non-linear optimization. It accelerates kNN operations with spherical bin partitioning optimized for the distribution of LiDAR data to eliminate unnecessary search spaces. In addition, the proposed spatio-temporal-aware computing minimizes excessive memory overhead and workload imbalance in kNN and PNN operations. Consequently, fabricated with 28-nm CMOS technology, the processor achieves 8.245 mJ/frame of energy consumption and a maximum performance of 20.7 ms latency, successfully demonstrating real-time semantic LiDAR SLAM system with 99.86% lower power consumption compared to modern CPU+GPU platforms.
- 6 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 60 (1) , 112-124
An Energy-Efficient Processor for Real-Time Semantic LiDAR SLAM in Mobile Robots
Emerging mobile robots require simultaneous localization and mapping (SLAM) systems with advanced 3-D perception and long-range 360 $^{\circ}$ interaction for autonomous driving. However, previous SLAM processors, which targeted only camera-based visual SLAM, are unsuitable for autonomous driving systems due to their limited field of view (FoV), inaccurate depth estimation, and lack of perception. In contrast, LiDAR offers a long-range 3-D point cloud with precise depth information and 360 $^{\circ}$ FoV, enabling the capture of fine environmental details. With its high accuracy and environmental robustness, LiDAR SLAM with accurate 3-D perception, semantic LiDAR SLAM, emerges as the most promising solution in autonomous driving systems. Nevertheless, real-time system-on-chip (SoC) implementation of semantic LiDAR SLAM has not been reported, primarily due to memory-intensive and compute-intensive operations caused by the simultaneous execution of multiple algorithms. Moreover, achieving real-time performance has not been feasible, even in the high-performance CPU $+$ GPU. In this article, a real-time and fully integrated semantic LiDAR SLAM processor (LSPU) is presented with semantic LiDAR-PNN-SLAM (LP-SLAM) system, which provides point neural network (PNN)-based 3-D segmentation, localization, and mapping simultaneously. The LSPU executes the LP-SLAM with the following features: 1) a $k$ -nearest neighbor (kNN) cluster with 2-D/3-D spherical coordinate-based bin (SB) searching to eliminate external memory access through dynamic memory allocation; 2) a PNN engine (PNNE) with a global point-level task scheduler (GPTS) to maximize core utilization by two-step workload balancing; 3) a keypoint extraction core (KEC) to skip redundant computation in the sorting operation; and 4) an optimization cluster with reconfigurable computation modes to support keypoint-level pipelining and parallel processing in non-linear optimization (NLO). As a result, the proposed LSPU achieves 20.7 ms of processing time, demonstrating real-time semantic LP-SLAM while consuming 99.89% lower energy compared to modern CPU $+$ GPU platforms.
- 6 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 60 (1) , 112-124
An Energy-Efficient Processor for Real-Time Semantic LiDAR SLAM in Mobile Robots
Emerging mobile robots require simultaneous localization and mapping (SLAM) systems with advanced 3-D perception and long-range 360 $^{\circ}$ interaction for autonomous driving. However, previous SLAM processors, which targeted only camera-based visual SLAM, are unsuitable for autonomous driving systems due to their limited field of view (FoV), inaccurate depth estimation, and lack of perception. In contrast, LiDAR offers a long-range 3-D point cloud with precise depth information and 360 $^{\circ}$ FoV, enabling the capture of fine environmental details. With its high accuracy and environmental robustness, LiDAR SLAM with accurate 3-D perception, semantic LiDAR SLAM, emerges as the most promising solution in autonomous driving systems. Nevertheless, real-time system-on-chip (SoC) implementation of semantic LiDAR SLAM has not been reported, primarily due to memory-intensive and compute-intensive operations caused by the simultaneous execution of multiple algorithms. Moreover, achieving real-time performance has not been feasible, even in the high-performance CPU $+$ GPU. In this article, a real-time and fully integrated semantic LiDAR SLAM processor (LSPU) is presented with semantic LiDAR-PNN-SLAM (LP-SLAM) system, which provides point neural network (PNN)-based 3-D segmentation, localization, and mapping simultaneously. The LSPU executes the LP-SLAM with the following features: 1) a $k$ -nearest neighbor (kNN) cluster with 2-D/3-D spherical coordinate-based bin (SB) searching to eliminate external memory access through dynamic memory allocation; 2) a PNN engine (PNNE) with a global point-level task scheduler (GPTS) to maximize core utilization by two-step workload balancing; 3) a keypoint extraction core (KEC) to skip redundant computation in the sorting operation; and 4) an optimization cluster with reconfigurable computation modes to support keypoint-level pipelining and parallel processing in non-linear optimization (NLO). As a result, the proposed LSPU achieves 20.7 ms of processing time, demonstrating real-time semantic LP-SLAM while consuming 99.89% lower energy compared to modern CPU $+$ GPU platforms.