• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.47 no.4
• /
• pp.82-87
• /
• 2010

Modern GPUs(Graphic Processing Units) provide computing capability higher than that of the general CPUs(Central Processor Units). With supports of programmability of graphics pipeline GP-GPU(General Purpose computation on GPU) has gained much attention expanding its application area. This paper compares sequential and massively parallel implementations of FDTD(Finite Difference Time Domain) algorithm using CUDA(Compute Unified Device Architecture). Experimental results show upto 45X speedup over conventional CPU execution.

• Journal of IKEEE
• /
• v.23 no.4
• /
• pp.1314-1320
• /
• 2019

In this paper, we propose a new method for optimizing bit operations and applying them to DNN(Deep Neural Network) in software environment. As a method for this, we propose a packing function for bitwise optimization and a masking matrix multiplication operation for application to DNN. The packing function converts 32-bit real value to 2-bit quantization value through threshold comparison operation. When this sequence is over, four 32-bit real values are changed to one 8-bit value. The masking matrix multiplication operation consists of a special operation for multiplying the packed weight value with the normal input value. And each operation was then processed in parallel using a GPU accelerator. As a result of this experiment, memory saved about 16 times than 32-bit DNN Model. Nevertheless, the accuracy was within 1%, similar to the 32-bit model.

• Journal of the Korea Computer Graphics Society
• /
• v.25 no.3
• /
• pp.105-113
• /
• 2019

Physically-based fluid simulation takes a lot of time for high resolution. To solve this problem, there are studies that make up the limitation of low resolution fluid simulation by using deep running. Among them, Super-resolution, which converts low-resolution simulation data to high resolution is under way. However, traditional techniques require to the entire space where there are no density data, so there are problems that are inefficient in terms of the full simulation speed and that cannot be computed with the lack of GPU memory as input resolution increases. In this paper, we propose a new method that divides and classifies 2D smoke simulation data into the space using the quad tree, one of the spatial partitioning methods, and performs Super-resolution only required space. This technique accelerates the simulation speed by computing only necessary space. It also processes the divided input data, which can solve GPU memory problems.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.19 no.4
• /
• pp.119-128
• /
• 2019

Due to the nature of the 3D graphics processor system, many mathematical calculations are required and parallel processing research using GPU (Graphics Processing Unit) is being performed for high-speed processing. In this paper, we propose a 3D graphics processor using MAMS, a parallel processor that does not use cache memory, to solve the GPU problem of increasing bandwidth caused by cache memory miss and the problem that 3D shader processing speed is not constant. The 3D graphics processor using MAMS proposed in this paper designed Vertex shader, Pixel shader, Tiling and Rasterizing structure using DirectX command analysis, the FPGA(Xilinx Virtex6@100MHz) board for MAMS was constructed and designed using Verilog. We compared the processing time of the developed FPGA (100Mhz) and nVidia GeForce GTX 660 (980Mhz), the processing time using GTX 660 was not constant and suing MAMS was constant.

• Journal of Korea Multimedia Society
• /
• v.12 no.2
• /
• pp.209-218
• /
• 2009

3D backprojection is a kind of reconstruction algorithm to generate volume data consisting of tomographic images, which provides spatial information of the original 3D data from hundreds of 2D projections. The computational time of backprojection increases in proportion to the size of volume data and the number of projection images since the value of every voxel in volume data is calculated by considering corresponding pixels from hundreds of projections. For the reduction of computational time, fast GPU based 3D backprojection methods have been studied recently and the performance of them has been improved significantly. This paper presents two multiple GPU based methods to maximize the parallelism of GPU and compares the efficiencies of two methods by considering both the number of projections and the size of volume data. The first method is to generate partial volume data independently for all projections after allocating a half size of volume data on each GPU. The second method is to acquire the entire volume data by merging the incomplete volume data of each GPU on CPU. The in-complete volume data is generated using the half size of projections after allocating the full size of volume data on each GPU. In experimental results, the first method performed better than the second method when the entire volume data can be allocated on GPU. Otherwise, the second method was efficient than the first one.

• Annual Conference of KIPS
• /
• 2023.11a
• /
• pp.79-80
• /
• 2023

동기식 분산 딥러닝 기법은 그래디언트 계산 작업을 다수의 워커가 나누어 병렬 처리함으로써 모델 학습 과정을 효율적으로 단축시킨다. 배치 사이즈는 이터레이션 단위로 처리하는 데이터 개수를 의미하며, 학습 속도 및 학습 모델의 품질에 영향을 미치는 중요한 요소이다. 멀티 GPU 환경에서 작동하는 분산 학습의 경우, 가용 GPU 메모리 용량이 커짐에 따라 선택 가능한 배치 사이즈의 상한이 증가한다. 하지만 배치 사이즈가 학습 속도 및 학습 모델 품질에 미치는 영향은 GPU 활용률, 총 에포크 수, 모델 파라미터 개수 등 다양한 변수에 영향을 받으므로 최적값을 찾기 쉽지 않다. 본 연구는 동기식 분산 딥러닝 환경에서 실험을 통해 최적의 배치 사이즈 선택에 영향을 미치는 주요 요인을 분석한다.

• The Journal of Korean Institute of Next Generation Computing
• /
• v.15 no.2
• /
• pp.39-49
• /
• 2019

Recently, CPU-GPU integrated heterogeneous multicore processors have been widely used for improving the performance of computing systems. Heterogeneous multicore processors integrate CPUs and GPUs on a single chip where CPUs and GPUs share the LLC(Last Level Cache). This causes a serious cache contention problem inside the processor, resulting in significant performance degradation. In this paper, we propose the partitioned LLC architecture to solve the cache contention problem in heterogeneous multicore processors. We analyze the performance impact varying the LLC size of CPUs and GPUs, respectively. According to our simulation results, the bigger the LLC size of the CPU, the CPU performance improves by up to 21%. However, the GPU shows negligible performance difference when the assigned LLC size increases. In other words, the GPU is less likely to lose the performance when the LLC size decreases. Because the performance degradation due to the LLC size reduction in GPU is much smaller than the performance improvement due to the increase of the LLC size of the CPU, the overall performance of heterogeneous multicore processors is expected to be improved by applying partitioned LLC to CPUs and GPUs. In addition, if we develop a memory management technique that can maximize the performance of each core in the future, we can greatly improve the performance of heterogeneous multicore processors.

• KIPS Transactions on Computer and Communication Systems
• /
• v.4 no.8
• /
• pp.239-244
• /
• 2015

WebCL was proposed for high complex computing in Javascript. Since WebCL-based applications are distributed and executed on an unspecified number of general clients, it is important to profile their performances on different clients. Several profilers have been introduced to support various programming languages but WebCL profiler has not been developed yet. In this paper, we present a WebCL profiler to evaluate WebCL-based applications and monitor the status of GPU on which they run. This profiler helps developers know the execution time of applications, memory read/write time, GPU statues such as its power consumption, temperature, and clock speed.

• Annual Conference of KIPS
• /
• 2016.10a
• /
• pp.62-63
• /
• 2016

딥러닝의 대표적 개발 환경 중 하나인 Caffe를 임베디드 시스템의 메모리 구조를 고려하여 최적화하고 실제 측정 실험으로 기존의 방식보다 처리시간과 소비 전력량의 이득이 있다는 것을 확인하였다. 구체적으로 통합 메모리를 사용하는 임베디드 시스템 환경의 특성에 적합한 zero-copy기법을 적용하여 CPU와 GPU 모두 접근이 가능하도록 메모리 영역을 맵핑하는 방식으로 메모리 복제에 따른 오버헤드를 줄였으며, GoogLeNet 네트워크 모델에 대하여 10%의 처리 속도 향상과, 36% 소비 전력 감소를 확인하였다.

• Journal of the Korea Computer Graphics Society
• /
• v.17 no.3
• /
• pp.1-7
• /
• 2011

Recently, various researches have been proposed to accelerate GPU-based volume ray-casting. However, those researches may cause several problems such as bottleneck of data transmission between CPU and GPU, requirement of additional video memory for hierarchical structure and increase of processing time whenever opacity transfer function changes. In this paper, we propose an efficient GPU-based empty space skipping technique to solve these problems. We store maximum density in a brick of volume dataset on a vertex element. Then we delete vertices regarded as transparent one by opacity transfer function in geometry shader. Remaining vertices are used to generate bounding boxes of non-transparent area that helps the ray to traverse efficiently. Although these vertices are independent on viewing condition they need to be reproduced when opacity transfer function changes. Our technique provides fast generation of opaque vertices for interactive processing since the generation stage of the opaque vertices is running in GPU pipeline. The rendering results of our algorithm are identical to the that of general GPU ray-casting, but the performance can be up to more than 10 times faster.