• Title/Summary/Keyword: Lee Taeho

Search Result 138, Processing Time 0.026 seconds

Optimizing 2-stage Tiling-based Matrix Multiplication in FPGA-based Neural Network Accelerator (FPGA기반 뉴럴네트워크 가속기에서 2차 타일링 기반 행렬 곱셈 최적화)

  • Jinse, Kwon;Jemin, Lee;Yongin, Kwon;Jeman, Park;Misun, Yu;Taeho, Kim;Hyungshin, Kim
    • IEMEK Journal of Embedded Systems and Applications
    • /
    • v.17 no.6
    • /
    • pp.367-374
    • /
    • 2022
  • The acceleration of neural networks has become an important topic in the field of computer vision. An accelerator is absolutely necessary for accelerating the lightweight model. Most accelerator-supported operators focused on direct convolution operations. If the accelerator does not provide GEMM operation, it is mostly replaced by CPU operation. In this paper, we proposed an optimization technique for 2-stage tiling-based GEMM routines on VTA. We improved performance of the matrix multiplication routine by maximizing the reusability of the input matrix and optimizing the operation pipelining. In addition, we applied the proposed technique to the DarkNet framework to check the performance improvement of the matrix multiplication routine. The proposed GEMM method showed a performance improvement of more than 2.4 times compared to the non-optimized GEMM method. The inference performance of our DarkNet framework has also improved by at least 2.3 times.

PartitionTuner: An operator scheduler for deep-learning compilers supporting multiple heterogeneous processing units

  • Misun Yu;Yongin Kwon;Jemin Lee;Jeman Park;Junmo Park;Taeho Kim
    • ETRI Journal
    • /
    • v.45 no.2
    • /
    • pp.318-328
    • /
    • 2023
  • Recently, embedded systems, such as mobile platforms, have multiple processing units that can operate in parallel, such as centralized processing units (CPUs) and neural processing units (NPUs). We can use deep-learning compilers to generate machine code optimized for these embedded systems from a deep neural network (DNN). However, the deep-learning compilers proposed so far generate codes that sequentially execute DNN operators on a single processing unit or parallel codes for graphic processing units (GPUs). In this study, we propose PartitionTuner, an operator scheduler for deep-learning compilers that supports multiple heterogeneous PUs including CPUs and NPUs. PartitionTuner can generate an operator-scheduling plan that uses all available PUs simultaneously to minimize overall DNN inference time. Operator scheduling is based on the analysis of DNN architecture and the performance profiles of individual and group operators measured on heterogeneous processing units. By the experiments for seven DNNs, PartitionTuner generates scheduling plans that perform 5.03% better than a static type-based operator-scheduling technique for SqueezeNet. In addition, PartitionTuner outperforms recent profiling-based operator-scheduling techniques for ResNet50, ResNet18, and SqueezeNet by 7.18%, 5.36%, and 2.73%, respectively.

Wafer TTV Measurement and Variable Effect Analysis According to Settling Time (Settling Time에 따른 웨이퍼 TTV 측정 및 변수 영향 분석)

  • Hyeong Won Kim;Anmok Jeong;Taeho Kim;Hak Jun Lee
    • Journal of the Semiconductor & Display Technology
    • /
    • v.22 no.3
    • /
    • pp.8-13
    • /
    • 2023
  • High bandwidth memory a core technology of the future memory semiconductor industry, is attracting attention. Temporary bonding and debonding process technology, which plays an important role in high bandwidth memory process technology, is also being studied. In this process, total thickness variation is a major factor determining wafer performance. In this study, the reliability of the equipment measuring total thickness variation is identified, and the servo motor settling, and wafer total thickness variation measurement accuracy are analyzed. As for the experimental variables, vacuum, acceleration time, and speed are changed to find the most efficient value by comparing the stabilization time. The smaller the vacuum and the larger the radius, the longer the settling time. If the radius is small, high-speed rotation performance is good, and if the radius is large, low-speed rotation performance is good. In the future, we plan to conduct an experiment to measure the entire of the wafer.

  • PDF

Discovery of a Novel 2,6-Difunctionalized 2H-Benzopyran Inhibitors Toward Sphingosylphosphorylcholine Synthetic Pathway as New Anti-inflammatory Target

  • Lee, Gee-Hyung;Lee, Seong Jin;Jeong, Dae Young;Kim, Ha-Young;Lee, Doohyun;Lee, Taeho;Hwang, Jong-Yeon;Park, Woo Kyu;Kong, Jae-Yang;Cho, Heeyeong;Gong, Young-Dae
    • Bulletin of the Korean Chemical Society
    • /
    • v.35 no.8
    • /
    • pp.2385-2390
    • /
    • 2014
  • Novel 2,6-difuctionalized 2H-benzopyrans were synthesized and evaluated for a sphingosylphosphorylcholine(SPC) inhibitor. The synthetic 2H-benzopyrans 1c and 3a showed high potency in SPC-induced cell proliferation assay ($IC_{50}$ < 20 nM). Neither hERG $K^+$ channel binding (> $10{\mu}M$) nor CYP inhibitions (> $10{\mu}M$) were observed. Also, the simple structure-activity relationship (SAR) results were obtained from analysis of 2H-benzopyran derivatives 1-3 and the anti-SPC effect of 2H-benzopyran 1c was confirmed by a HUVEC tube formation assay.

Identification of ML106 Phase 1 Metabolites in Human Liver Microsomes Using High-Resolution Quadrupole-Orbitrap Mass Spectrometry

  • Jo, Jun Hyeon;Nam, WoongShik;Kim, Sunjoo;Lee, Doohyun;Min, Kyung Hoon;Lee, Taeho;Lee, Sangkyu
    • Mass Spectrometry Letters
    • /
    • v.7 no.3
    • /
    • pp.69-73
    • /
    • 2016
  • High-resolution quadrupole-Orbitrap mass spectrometry (HRMS), with high-resolution (> 10,000 at full-width at half-maximum) and accurate mass (< 5 ppm deviation) capabilities, plays an important role in the structural elucidation of drug metabolites in the pharmaceutical industry. ML106, a derivative of imidazobenzimidazole, decreased melanin content and tyrosinase activity in a dose-dependent manner. Here, we investigated the phase 1 metabolic pathway of ML106 using HRMS in human liver microsomes (HLMs) and recombinant cDNA-expressed cytochrome P450 (CYP). After the incubation of ML106 with pooled HLMs and recombinant cDNA-expressed CYP in the presence of NADPH, five phase 1 metabolites, including three mono-hydroxylated metabolites (M1-3) and two di-hydroxylated metabolites (M4 and M5), were investigated. The metabolite structures were postulated by the elucidation of protonated mass spectra using HRMS. The CYP isoforms related to the hydroxylation of ML106 were studied after incubation with recombinant cDNA-expressed CYP. Here, we identified the phase 1 metabolic pathway of ML106 induced by CYP in HLMs.

The First Report to Evaluate Safety of Cyanobacterium Leptolyngbya sp. KIOST-1 for Use as a Food Ingredient: Oral Acute Toxicity and Genotoxicity Study

  • Lee, Youngdeuk;Kim, Taeho;Lee, Won-Kyu;Ryu, Yong-Kyun;Kim, Ji Hyung;Jeong, Younsik;Park, Areumi;Lee, Yeon-Ji;Oh, Chulhong;Kang, Do-Hyung
    • Journal of Microbiology and Biotechnology
    • /
    • v.31 no.2
    • /
    • pp.290-297
    • /
    • 2021
  • Leptolyngbya sp. KIOST-1 (LK1) is a newly isolated cyanobacterium that shows no obvious cytotoxicity and contains high protein content for both human and animal diets. However, only limited information is available on its toxic effects. The purpose of this study was to validate the safety of LK1 powder. Following Organisation for Economic Co-operation and Development (OECD) guidelines, a single-dose oral toxicity test in Sprague Dawley rats was performed. Genotoxicity was assessed using a bacterial reverse mutation test with Salmonella typhimurium (strains TA98, TA100, TA1535, and TA1537) and Escherichia coli WP2 uvrA, an in vitro mammalian chromosome aberration test using Chinese hamster lung cells, and an in vivo mammalian erythrocyte micronucleus test using Hsd:ICR (CD-1) SPF mouse bone marrow. After LK1 administration (2,500 mg/kg), there were no LK1-related body weight changes or necropsy findings. The reverse mutation test showed no increased reverse mutation upon exposure to 5,000 ㎍/plate of the LK1 powder, the maximum tested amount. The chromosome aberration test and micronucleus assay demonstrated no chromosomal abnormalities and genotoxicity, respectively, in the presence of the LK1 powder. The absence of physiological findings and genetic abnormalities suggests that LK1 powder is appropriate as a candidate biomass to be used as a safe food ingredient.

Complete Chloroplast Genome assembly and Annotation of Milk Thistle (Silybum marianum) and Phylogenetic Analysis

  • Hwajin Jung;Yedomon Ange Bovys Zoclanclounon;Jeongwoo Lee;Taeho Lee;Jeonggu Kim;Guhwang Park;Keunpyo Lee;Kwanghoon An;Jeehyoung Shim;Joonghyoun Chin;Suyoung Hong
    • Proceedings of the Korean Society of Crop Science Conference
    • /
    • 2022.10a
    • /
    • pp.210-210
    • /
    • 2022
  • Silybum marianum is an annual or biennial plant from the Asteraceae family. It can grow in low-nutrient soil and drought conditions, making it easy to cultivate. From the seed, a specialized plant metabolite called silymarin (flavonolignan complex) is produced and is known to alleviate the liver from hepatitis and toxins damages. To infer the phylogenetic placement of a Korean milk thistle, we conducted a chloroplast assembly and annotation following by a comparison with existing Chinese reference genome (NC_028027). The chloroplast genome structure was highly similar with an assembly size of 152,642 bp, an 153,202 bp for Korean and Chinese milk thistle respectively. Moreover, there were similarities at the gene level, coding sequence (n = 82), transfer RNA (n = 31) and ribosomal RNA (n = 4). From all coding sequences gene set, the phylogenetic tree inference placed the Korean cultivar into the milk thistle clade; corroborating the expected tree. Moreover, an investigation the tree based only on the ycf1 gene confirmed the same tree; suggesting that ycf1 gene is a potential marker for DNA barcoding and population diversity study in milk thistle genus. Overall, the provided data represents a valuable resource for population genomics and species-centered determination since several species have been reported in the Silybum genus.

  • PDF

A Rapid and Efficient Screening Method for Antibacterial Compound-Producing Bacteria

  • Hettiarachchi, Sachithra Amarin;Lee, Su-Jin;Lee, Youngdeuk;Kwon, Young-Kyung;Zoysa, Mahanama De;Moon, Song;Jo, Eunyoung;Kim, Taeho;Kang, Do-Hyung;Heo, Soo-Jin;Oh, Chulhong
    • Journal of Microbiology and Biotechnology
    • /
    • v.27 no.8
    • /
    • pp.1441-1448
    • /
    • 2017
  • Antibacterial compounds are widely used in the treatment of human and animal diseases. The overuse of antibiotics has led to a rapid rise in the prevalence of drug-resistant bacteria, making the development of new antibacterial compounds essential. This study focused on developing a fast and easy method for identifying marine bacteria that produce antibiotic compounds. Eight randomly selected marine target bacterial species (Agrococcus terreus, Bacillus algicola, Mesoflavibacter zeaxanthinifaciens, Pseudoalteromonas flavipulchra, P. peptidolytica, P. piscicida, P. rubra, and Zunongwangia atlantica) were tested for production of antibacterial compounds against four strains of test bacteria (B. cereus, B. subtilis, Halomonas smyrnensis, and Vibrio alginolyticus). Colony picking was used as the primary screening method. Clear zones were observed around colonies of P. flavipulchra, P. peptidolytica, P. piscicida, and P. rubra tested against B. cereus, B. subtilis, and H. smyrnensis. The efficiency of colony scraping and broth culture methods for antimicrobial compound extraction was also compared using a disk diffusion assay. P. peptidolytica, P. piscicida, and P. rubra showed antagonistic activity against H. smyrnensis, B. cereus, and B. subtilis, respectively, only in the colony scraping method. Our results show that colony picking and colony scraping are effective, quick, and easy methods of screening for antibacterial compound-producing bacteria.

Optical Diagnostic Study for Flame Characteristic Analysis in Aluminum Dust Clouds (알루미늄 군입자 화염특성 분석을 위한 광학기법 연구)

  • Lee, Sanghyup;Ko, Taeho;Lim, Jihwan;Lee, Dohyung;Yoon, Woongsup
    • Journal of the Korean Society of Propulsion Engineers
    • /
    • v.17 no.5
    • /
    • pp.47-53
    • /
    • 2013
  • In this study, In order to develop the measurement method of high energy density metal aluminum dust cloud combustion, flame temperature and emission spectrum was measured using spectrometer. Because of the ultra high ${\mu}m$-sized aluminum flame temperature more than 2400 K, it was measured by non-contact optical technique which is the modified two wavelength pyrometry with 520, 640 nm and spectrum comparison method. These methods were applied to experiment after accurate verification. As a result, we could identify that flame temperature is more than 2400 K in bottom of combustor in both methods. And on the emission spectrum analysis, we could measure AlO radical which is occurred dominantly in aluminum combustion.

FastIO: High Speed Launching of Smart TV Apps (FastIO: 스마트 TV 앱의 고속 구동 기법)

  • Lee, Cheolhee;Hwang, Taeho;Won, Youjip;Lee, Seongjin
    • Journal of KIISE
    • /
    • v.43 no.7
    • /
    • pp.725-735
    • /
    • 2016
  • Smart TV uses Webkit as a web browser engine to provide contents such as web surfing, VOD watching, and games. Webkit uses web resources, such as HTML, CSS, JavaScript, and images, in order to run applications. At the start of an application, Webkit loads resources to the memory and creates DOM tree and render tree, which is a time consuming process. However, DOM tree and render tree created by the smart TV application do not change over time because the smart TV application uses web resources stored in a disk. If DOM tree and render tree can be stored and reused, it is possible to reduce loading time of an application. In this paper, we propose FastIO technique that selectively adds persistency to dynamically allocated memory. FastIO reduces overall application loading time by eliminating the process of loading resources from storage, parsing the HTML documents, and creating DOM tree and render tree. Comparison of the application resource loading times indicates that the web browser with FastIO is 7.9x, 44.8x, and 2.9x faster than the legacy web browser in an SSD, Ramdisk, and eMMC environment, respectively.