https://scholar.dgist.ac.kr/handle/20.500.11750/371 2026-07-21T11:32:19Z https://scholar.dgist.ac.kr/handle/20.500.11750/60457 Title: MMC: Metadata Migration for Efficient Memory Management in CXL DRAM Systems Author(s): Lee, Dusol; Lee, Taehyung; Lee, Hanseon; Kim, Nam Sung; Kim, Jihong Abstract: CXL-DRAM enables cost-efficient memory expansion in data-center servers, but introduces new challenges for operating system memory management. In particular, large in-kernel metadata structures can occupy a significant portion of DDR-DRAM, reducing the space available for application data and increasing accesses to slower CXL-DRAM. Placing all metadata in CXL-DRAM frees DDR-DRAM space but can cause severe tail latency due to inefficient memory reclamation. We propose MMC, a hot-cold metadata migration mechanism that balances application performance and reclamation efficiency by dynamically placing metadata across DDR-DRAM and CXL-DRAM. Our results show that MMC improves throughput while reducing tail latency across diverse workloads. 2025-12-31T15:00:00Z https://scholar.dgist.ac.kr/handle/20.500.11750/60450 Title: A depth-customizable double-sided 3D neural probe array for simultaneous investigation of multiple brain regions Author(s): Park, Sehwan; Kim, Minseok; Lee, Haeyun; Lee, Jimin; Chou, Namsun; Shin, Hyogeun Abstract: Understanding the complex neural circuits within the brain requires advanced tools capable of simultaneously recording signals from multiple regions and depths. However, previously developed tools have limited capability to address 3D structures in the brain as they typically feature fixed probe lengths and single-sided electrode configurations. To overcome these challenges, we developed a depth-customizable 3D electrode array structure comprising double-sided 2D neural probe arrays via flexible printed circuit board technology with a zeroinsertion-force connector and a supporting board without requiring additional fabrication steps. This enables precise depth adjustments and the double-sided electrode configuration effectively doubles the number of recording sites, thereby facilitating volumetric and comprehensive neural signal acquisition. Our device allows user-defined adjustment of probe spacing, achieving a minimum inter-probe distance of 1 mm, and enables finetuned control of insertion depth for precise targeting of specific brain regions, with a maximum depth difference of only 0.168 mm. Also, by employing a PSR ink insulation layer, we achieved a total probe thickness of approximately 80 mu m, resulting in a compact design that eliminates the need for complex semiconductor processes. Validation of the device in vivo demonstrated its capability to simultaneously monitor neural signals from multiple brain regions. Its depth-customizable design facilitated functional connectivity studies across various depths, the results of which could provide critical insights into neural network dynamics. Our approach significantly enhances the flexibility, scalability, and efficiency of neural probes and provides a powerful platform for neuroscience research into areas such as brain-machine interface development and functional connectivity. 2025-11-30T15:00:00Z https://scholar.dgist.ac.kr/handle/20.500.11750/60449 Title: Atomic Layer Modulation of Ruthenium Aluminum Oxide through Reactivity Control of Precursors for Seedless Copper Interconnects Author(s): Nguyen, Chi Thang; Kim, Miso; Kim, Youn-Hye; Cheon, Taehoon; Trinh, Ngoc Le; Kim, Sehee; Kim, Soo-Hyun; Shong, Bonggeun; Lee, Han-Bo-Ram Abstract: We fabricated the RuAlO x multicomponent thin film by using atomic layer modulation (ALM) based on precursor chemical reactivities and steric hindrance effects. Dicarbonyl-bis(5-methyl-2,4-hexanediketonato)Ru(II) (Carish) and dimethylaluminum isopropoxide (DMAI) were employed as Ru and Al precursors, respectively, with H2O as the counter reactant. Theoretical calculations based on machine learning interatomic potential were performed to investigate the surface chemical reactions of the precursors and the feasibility of the ALM concept to modulate RuAlO x films. The transmission electron microscopy analysis revealed a distinctive structure of the RuAlO x thin films, where the typical columnar growth of Ru was prevented by the surrounding amorphous Al2O3. Sheet resistance measurement results and X-ray diffraction analyses confirmed that a 50 nm Cu/5 nm RuAlO x /SiO2 structure remained stable even after annealing at 600 degrees C for 30 min, without any Cu silicide formation. These results suggest that a 5 nm RuAlO x thin film effectively prevents the diffusion of 50 nm of Cu. We believe that RuAlO x ALM thin films can be used as diffusion barriers against Cu, with improved performance compared to that of films with the typical Ru columnar grain structure. 2025-11-30T15:00:00Z https://scholar.dgist.ac.kr/handle/20.500.11750/60447 Title: SearchLight: Neural Architecture Search for Lightweight Spatio-Temporal Graph Neural Networks Author(s): Yoon, Heeyong; Jung, Jinhong; Chon, Kang-Wook; Kim, Min-Soo Abstract: Spatio-Temporal Graph Neural Networks (STGNNs) are neural network models that integrate spatial information into time series processing, and have been successfully applied in various applications. Although these models have demonstrated strong prediction capabilities, most existing STGNN architectures require a significant memory size and long training times. Some lightweight versions of STGNNs have been proposed, but they still rely on expert-driven manual designs to improve performance, which require implicit domain-specific knowledge that varies across datasets. This design approach limits their adaptability to different application scenarios. To address this limitation, Neural Architecture Search (NAS) has been applied to automate the STGNN design process. However, existing NAS-based approaches prioritize prediction accuracy rather than resource efficiency. As a result, current approaches fail to provide compact model architectures or efficient training and limit their scalability. In this work, we introduce SearchLight, a novel NAS-based STGNN framework to automatically discover lightweight STGNN models while maintaining prediction performance. We set two cells for spatial and temporal operations into two distinct sets to capture spatial and temporal data features better for each cell type of the NAS method. We specialize in cell types for spatial and temporal information so that the model can better capture and combine the intrinsic features of spatial and temporal data. We employ a multi-objective search strategy that optimizes both model compactness and prediction accuracy to enable our method to discover lightweight and accurate STGNN models. Experimental results across several real-world datasets show that SearchLight reduces the model size by an average of and training time by an average of , while sacrificing a small amount of prediction performance, an average of 1.6%p, compared to manually designed and existing NAS-based STGNN models. 2025-08-31T15:00:00Z