Hoi-Jun Yoo
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Hoi-Jun Yoo has a total of 120 co-authors. Here are the top co-authors by number of shared publications:
- Donghyeon Han
Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology (MIT), Cambridge, MA, USA
37 shared publications - Sangjin Kim
Korea Advanced Institute of Science and Technology (KAIST)
34 shared publications - Sangyeob Kim
School of Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea
30 shared publications - Gwangtae Park
School of Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea
21 shared publications - Dongseok Im
School of Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea
16 shared publications - Juhyoung Lee
School of Electrical Engineering, Korea Advanced Institute of Science and Technology, Daejeon, South Korea
16 shared publications
Latest Publications
- Open Access
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
SliceMoE: Bit-Sliced Expert Caching under Miss-Rate Constraints for Efficient MoE Inference
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessSliceMoE: Bit-Sliced Expert Caching under Miss-Rate Constraints for Efficient MoE Inference
- 22 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. PP (00189200) , 1-13
MIDAS: An Energy-Efficient Microscaling Quantization-Based Digital CIM Accelerator With Spatio-Temporal Cyclic Alignment for Generative AI
Digital compute-in-memory (DCIM) and the microscaling (MX) data format have emerged as promising hardware and algorithmic solutions, respectively, for energy-efficient generative AI (GenAI) processing. However, their combination introduces challenges, including redundant computation from mantissa alignment, excessive exponent handling overhead, and accuracy loss from static approximation schemes that fail to accommodate block-wise dynamic bit-significance. In this article, we present MIDAS, an MX-optimized DCIM-based GenAI accelerator with three key features: 1) spatio-temporal cyclic alignment (STCA) with sign-magnitude-based assembled column addition (SACA) enables compact weight mapping, high input reusability, and low-latency computation, achieving $1.72\times $ higher macro energy efficiency; 2) MX-aware minimum exponent post-detection (MEPD) eliminates redundant maximum exponent detection by exploiting the inherent characteristics of MX quantization, reducing exponent handling power, area, and latency by 68.8%, 63.2%, and 22.5%, respectively; and 3) sense amplifier-based dynamic bit (SADB)-significance handling with shared scale-aware channel regrouping (SSCR) dynamically adjusts approximation bit-positions according to each block’s shared scale, improving energy efficiency by 27.3% while minimizing accuracy loss. MIDAS is fabricated in Samsung 28-nm CMOS technology and occupies a 7.56 mm2 die area. MIDAS achieves macro energy and area efficiencies of 96.3 TFLOPS/W and 0.69 TFLOPS/mm2, respectively, under MXFP4. Compared to state-of-the-art DCIMs, MIDAS demonstrates $1.58\times $ higher system energy efficiency and $3.32\times $ higher macro energy efficiency, validating its effectiveness as an MX-optimized AI accelerator for next-generation GenAI applications.
- 22 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. PP (00189200) , 1-13
MIDAS: An Energy-Efficient Microscaling Quantization-Based Digital CIM Accelerator With Spatio-Temporal Cyclic Alignment for Generative AI
Digital compute-in-memory (DCIM) and the microscaling (MX) data format have emerged as promising hardware and algorithmic solutions, respectively, for energy-efficient generative AI (GenAI) processing. However, their combination introduces challenges, including redundant computation from mantissa alignment, excessive exponent handling overhead, and accuracy loss from static approximation schemes that fail to accommodate block-wise dynamic bit-significance. In this article, we present MIDAS, an MX-optimized DCIM-based GenAI accelerator with three key features: 1) spatio-temporal cyclic alignment (STCA) with sign-magnitude-based assembled column addition (SACA) enables compact weight mapping, high input reusability, and low-latency computation, achieving $1.72\times $ higher macro energy efficiency; 2) MX-aware minimum exponent post-detection (MEPD) eliminates redundant maximum exponent detection by exploiting the inherent characteristics of MX quantization, reducing exponent handling power, area, and latency by 68.8%, 63.2%, and 22.5%, respectively; and 3) sense amplifier-based dynamic bit (SADB)-significance handling with shared scale-aware channel regrouping (SSCR) dynamically adjusts approximation bit-positions according to each block’s shared scale, improving energy efficiency by 27.3% while minimizing accuracy loss. MIDAS is fabricated in Samsung 28-nm CMOS technology and occupies a 7.56 mm2 die area. MIDAS achieves macro energy and area efficiencies of 96.3 TFLOPS/W and 0.69 TFLOPS/mm2, respectively, under MXFP4. Compared to state-of-the-art DCIMs, MIDAS demonstrates $1.58\times $ higher system energy efficiency and $3.32\times $ higher macro energy efficiency, validating its effectiveness as an MX-optimized AI accelerator for next-generation GenAI applications.
- 10 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. PP (00189200) , 1-14
Falcon: Event-Driven Hybrid Object Tracking SoC With Bit-Column-Similarity-Based MRAM PIM and Heterogeneous Read-Only-Once Attention Flow
Hybrid frame-event cameras offer high spatial fidelity and $\mu $ s-level temporal response, making them well suited for low-power, low-latency object tracking under motion blur and sparse activity. This work presents an event-driven vision SoC that integrates heterogeneous MRAM-based processing-in-memory (PIM), near-memory-computing (NMC), lookup table (LUT), and single-instruction multiple-data (SIMD) compute units for end-to-end real-time hybrid tracking. First, the CNN feature encoder is mapped onto a bit-column-similarity-optimized MRAM PIM with activation skipping and weight compression, achieving $2.44\times $ energy efficiency and 61.9% latency reduction. Second, attention-based tracking is executed using a read-once $Q$ / $K$ / $V$ dataflow combining column-priority NMC generation, mixed-product SIMD execution, and LUT-accelerated SoftMax, reducing buffer reads by 46% and attention latency by 30%. Third, a two-stage event-driven activation pipeline selectively enables computation only upon motion, cutting system power by 63.5% and overall latency by 68.8%. Fabricated in 28-nm FDSOI (12.96 mm2), the SoC delivers macro-level 750.18 TOPS/W for 1b–1b–3b MAC operations at 0.7 V, 50MHz, and system-level 16.34 TOPS/W for INT8 CNN + FP8 attention at 1.15 V, 200MHz, enabling real-time 50 frames/p tracking at 196.5-mW power consumption.
- 10 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. PP (00189200) , 1-14
Falcon: Event-Driven Hybrid Object Tracking SoC With Bit-Column-Similarity-Based MRAM PIM and Heterogeneous Read-Only-Once Attention Flow
Hybrid frame-event cameras offer high spatial fidelity and $\mu $ s-level temporal response, making them well suited for low-power, low-latency object tracking under motion blur and sparse activity. This work presents an event-driven vision SoC that integrates heterogeneous MRAM-based processing-in-memory (PIM), near-memory-computing (NMC), lookup table (LUT), and single-instruction multiple-data (SIMD) compute units for end-to-end real-time hybrid tracking. First, the CNN feature encoder is mapped onto a bit-column-similarity-optimized MRAM PIM with activation skipping and weight compression, achieving $2.44\times $ energy efficiency and 61.9% latency reduction. Second, attention-based tracking is executed using a read-once $Q$ / $K$ / $V$ dataflow combining column-priority NMC generation, mixed-product SIMD execution, and LUT-accelerated SoftMax, reducing buffer reads by 46% and attention latency by 30%. Third, a two-stage event-driven activation pipeline selectively enables computation only upon motion, cutting system power by 63.5% and overall latency by 68.8%. Fabricated in 28-nm FDSOI (12.96 mm2), the SoC delivers macro-level 750.18 TOPS/W for 1b–1b–3b MAC operations at 0.7 V, 50MHz, and system-level 16.34 TOPS/W for INT8 CNN + FP8 attention at 1.15 V, 200MHz, enabling real-time 50 frames/p tracking at 196.5-mW power consumption.
- 27 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 61 (6) , 3032-3044
Space-Mate: A 303.5-mW Real-Time Sparse Mixture-of-Experts-Based NeRF-SLAM Processor for Mobile Spatial Computing
Simultaneous localization and mapping (SLAM) provides crucial ego-pose information and 3-D maps of the user environment, which are fundamental to emerging mobile spatial computing devices. Dense 3-D mapping and accurate pose estimation are particularly necessary for applications like augmented reality (AR) and autonomous navigation. However, existing SLAM processors are typically limited to sparse 3-D maps due to constrained resources in mobile settings. This article presents space-mate, a low-power, real-time dense SLAM processor. At the algorithm level, a sparse mixture-of-experts (SMoE) based neural radiance field (NeRF) is introduced as a novel representation for dense 3-D mapping, optimized for mobile devices due to its reduced computational complexity and compact map size. Space-mate executes both tracking and mapping on a unified hardware with the following optimizations: 1) out-of-order SMoE (OoO-SMoE) router schedules MLP operations for multiple experts, enabling multi-batch processing and reducing data transactions; 2) heterogeneous coarse-grained sparse core (HCG-SC) accelerates various MLP training stages, including feed-forward (FF), back-propagation (BP), and weight-gradient (WG) calculation, enhancing energy efficiency; and 3) 2-D sampling unit (2D-SU) predicts and removes unnecessary mapping workload through loss prediction, significantly reducing energy per frame with minimal accuracy loss. Space-mate, fabricated in a 28-nm process, achieves 32.85 frames/s at 303.5-mW power consumption under a 125 MHz, 0.83 V operating condition. Compared to prior state-of-the-art SLAM ASICs, space-mate processes$30.72\times $ –$930.9\times $ more points per frame for tracking and$4.8\times $ –$7.2\times $ more points for mapping while consuming$2\times $ –$15.4\times $ less energy per pixel.- 27 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 61 (6) , 3032-3044
Space-Mate: A 303.5-mW Real-Time Sparse Mixture-of-Experts-Based NeRF-SLAM Processor for Mobile Spatial Computing
Simultaneous localization and mapping (SLAM) provides crucial ego-pose information and 3-D maps of the user environment, which are fundamental to emerging mobile spatial computing devices. Dense 3-D mapping and accurate pose estimation are particularly necessary for applications like augmented reality (AR) and autonomous navigation. However, existing SLAM processors are typically limited to sparse 3-D maps due to constrained resources in mobile settings. This article presents space-mate, a low-power, real-time dense SLAM processor. At the algorithm level, a sparse mixture-of-experts (SMoE) based neural radiance field (NeRF) is introduced as a novel representation for dense 3-D mapping, optimized for mobile devices due to its reduced computational complexity and compact map size. Space-mate executes both tracking and mapping on a unified hardware with the following optimizations: 1) out-of-order SMoE (OoO-SMoE) router schedules MLP operations for multiple experts, enabling multi-batch processing and reducing data transactions; 2) heterogeneous coarse-grained sparse core (HCG-SC) accelerates various MLP training stages, including feed-forward (FF), back-propagation (BP), and weight-gradient (WG) calculation, enhancing energy efficiency; and 3) 2-D sampling unit (2D-SU) predicts and removes unnecessary mapping workload through loss prediction, significantly reducing energy per frame with minimal accuracy loss. Space-mate, fabricated in a 28-nm process, achieves 32.85 frames/s at 303.5-mW power consumption under a 125 MHz, 0.83 V operating condition. Compared to prior state-of-the-art SLAM ASICs, space-mate processes$30.72\times $ –$930.9\times $ more points per frame for tracking and$4.8\times $ –$7.2\times $ more points for mapping while consuming$2\times $ –$15.4\times $ less energy per pixel.- 7 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 61 (1) , 183-196
IRIS: An Energy-Efficient Spatial Computing SoC for Real-Time Interactive Rendering and Modeling With Surface-Aware 3-D Gaussian Splatting
The 3-D Gaussian splatting (3DGS), based on a machine learning-driven radiance field technique, is rapidly emerging as a next-generation solution in 3-D graphics. Owing to its short modeling time, computational simplicity, and high rendering quality, it is expected to replace traditional 3-D graphics on edge devices. However, its substantial memory access requirements and the challenges of real-time modification present major barriers to deployment on mobile augmented/virtual reality (AR/VR) platforms. We present a spatial computing system-on-chip (SoC), IRIS, to enable real-time, spatially interactive rendering with surface-aware modelling (SAM). The architecture of IRIS introduces three key hardware features. First, the surface perception unit (SPU) reduces external memory access (EMA) and minimizes Gaussian sorting latency by clustering Gaussians based on surface alignment. Second, the reconfigurable ApproxLP (ALP)-based multiply-accumulator array (RAMA) achieves high computational efficiency and throughput during neural Gaussian feature generation and mesh deformation. Finally, the error direction cache (EDC) enables fast and efficient modeling by eliminating redundant backpropagation (BP) operations. As a result, IRIS is fabricated using a 28-nm process and occupies a die area of 20.25 mm2. When evaluated on the SyntheticNeRF dataset, IRIS achieves $34.3{\times }$ faster interactive rendering and consumes $300.0{\times }$ less energy per frame compared to an edge GPU, despite having $21.3{\times }$ lower external memory bandwidth.
- 7 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Solid-State Circuits
- Vol. 61 (1) , 183-196
IRIS: An Energy-Efficient Spatial Computing SoC for Real-Time Interactive Rendering and Modeling With Surface-Aware 3-D Gaussian Splatting
The 3-D Gaussian splatting (3DGS), based on a machine learning-driven radiance field technique, is rapidly emerging as a next-generation solution in 3-D graphics. Owing to its short modeling time, computational simplicity, and high rendering quality, it is expected to replace traditional 3-D graphics on edge devices. However, its substantial memory access requirements and the challenges of real-time modification present major barriers to deployment on mobile augmented/virtual reality (AR/VR) platforms. We present a spatial computing system-on-chip (SoC), IRIS, to enable real-time, spatially interactive rendering with surface-aware modelling (SAM). The architecture of IRIS introduces three key hardware features. First, the surface perception unit (SPU) reduces external memory access (EMA) and minimizes Gaussian sorting latency by clustering Gaussians based on surface alignment. Second, the reconfigurable ApproxLP (ALP)-based multiply-accumulator array (RAMA) achieves high computational efficiency and throughput during neural Gaussian feature generation and mesh deformation. Finally, the error direction cache (EDC) enables fast and efficient modeling by eliminating redundant backpropagation (BP) operations. As a result, IRIS is fabricated using a 28-nm process and occupies a die area of 20.25 mm2. When evaluated on the SyntheticNeRF dataset, IRIS achieves $34.3{\times }$ faster interactive rendering and consumes $300.0{\times }$ less energy per frame compared to an edge GPU, despite having $21.3{\times }$ lower external memory bandwidth.