• Journal of IKEEE
• /
• v.18 no.3
• /
• pp.392-397
• /
• 2014

In this paper, we propose a memory operation system architecture for memory latency penalty reduction in SIMT architecture based stream processor. The proposed architecture applied non-blocking cache architecture to reduce cache miss penalty generated by blocking cache architecture. We verified that the proposed memory operation architecture improve the performance of the stream processor by comparing processing performances of various algorithms. We measured the performance improvement rate that was improved in accordance with the ratio of memory instruction in each algorithm. As a result, we confirmed that the performance of stream processor improves up to minimum 8.2% and maximum 46.5%.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.11 no.2
• /
• pp.57-65
• /
• 2016

In this paper, we propose a system for efficient use of shared memory between CPU and GPU. The system, called Fusion Architecture, assures consistency of the shared memory and minimizes cache misses that frequently occurs on Heterogeneous System Architecture or Unified Virtual Memory based systems. It also maximizes the performance for memory intensive jobs by efficient allocation of GPU cores. To test between architectures on various scenarios, we introduce the Fusion Architecture Analyzer, which compares OpenMP, OpenCL, CUDA, and the proposed architecture in terms of memory overhead and process time. As a result, Proposed fusion architectures show that the Fusion Architecture runs benchmarks 55% faster and reduces memory overheads by 220% in average.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.5 no.2
• /
• pp.77-83
• /
• 2010

In this paper, we propose an efficient memory architecture for coeff_token encoding in H.264/AVC standard. The VLCTs used to encode the coeff_token syntax element are implemented with the memory. In general, the size of memory must be reduced because it affects the cost and operation speed of the system. Based on the analysis for the codewords in VLCTs, new memory architecture is designed in this paper. The proposed memory architecture results in about 24% memory saving, compared to the conventional memory architecture.

• Journal of the Korean Society of Industry Convergence
• /
• v.24 no.1
• /
• pp.69-77
• /
• 2021

The widely used low-cost design methodology for IoT devices is very popular. In such a networked device, memory is composed of flash memory, SRAM, DRAM, etc., and because it processes a large amount of data, memory design is an important factor for system performance. Therefore, each device selects optimized design factors such as function, performance and cost according to market demand. The design of a memory architecture available for low-cost IoT devices is very limited with the configuration of SRAM, flash memory, and DRAM. In order to process as much data as possible in the same space, an architecture that supports parallel processing units is usually provided. Such parallel architecture is a design method that provides high performance at low cost. However, it needs precise software techniques for instruction and data mapping on the parallel architecture. This paper proposes an instruction/data mapping method to support optimized parallel processing performance. The proposed method optimizes system performance by actively using hardware and software parallelism.

• Journal of the Korea Society of Computer and Information
• /
• v.17 no.7
• /
• pp.1-11
• /
• 2012

As process technology scales down, the number of cores integrated into a processor increases dramatically, leading to significant performance improvement. Especially, the GPU(Graphics Processing Unit) containing many cores can provide high computational performance by maximizing the parallelism. In the GPU architecture, the access latency to the main memory becomes one of the major reasons restricting the performance improvement. In this work, we analyze the performance improvement of the 3D GPU architecture compared to the 2D GPU architecture quantitatively and investigate the potential problems of the 3D GPU architecture. In general, memory instructions account for 30% of total instructions, and global/local memory instructions constitutes 60% of total memory instructions. Therefore, the performance of the 3D GPU is expected to be improved significantly compared to the 2D GPU by reducing the delay of memory instructions. However, according to our experimental results, the 3D architecture improves the GPU performance only by 2% compared to the 2D architecture due to the memory bottleneck, since the performance reduction due to memory bottleneck in the 3D GPU architecture increases by 245% compared to the 2D architecture. This paper provides the guideline for suitable memory design by analyzing the efficiency of the memory architecture in 3D GPU architecture.

• ETRI Journal
• /
• v.42 no.4
• /
• pp.505-517
• /
• 2020

The increasing size and complexity of deep neural networks (DNNs) necessitate the development of efficient high-performance accelerators. An efficient memory structure and operating scheme provide an intuitive solution for high-performance accelerators along with dataflow control. Furthermore, the processing of various neural networks (NNs) requires a flexible memory architecture, programmable control scheme, and automated optimizations. We first propose an efficient architecture with flexibility while operating at a high frequency despite the large memory and PE-array sizes. We then improve the efficiency and usability of our architecture by automating the optimization algorithm. The experimental results show that the architecture increases the data reuse; a diagonal write path improves the performance by 1.44× on average across a wide range of NNs. The automated optimizations significantly enhance the performance from 3.8× to 14.79× and further provide usability. Therefore, automating the optimization as well as designing an efficient architecture is critical to realizing high-performance DNN accelerators.

• Electronics and Telecommunications Trends
• /
• v.39 no.1
• /
• pp.62-73
• /
• 2024

Artificial intelligence relies on data-driven analysis, and the data processing performance strongly depends on factors such as memory capacity, bandwidth, and latency. Fast and large-capacity memory can be achieved by composing numerous high-performance memory units connected via high-performance interconnects, such as Compute Express Link (CXL). CXL is designed to enable efficient communication between central processing units, memory, accelerators, storage, and other computing resources. By adopting CXL, a composable computing architecture can be implemented, enabling flexible server resource configuration using a pool of computing resources. Thus, manufacturers are actively developing hardware and software solutions to support CXL. We present a survey of the latest software for CXL memory utilization and the most recent CXL memory emulation software. The former supports efficient use of CXL memory, and the latter offers a development environment that allows developers to optimize their software for the hardware architecture before commercial release of CXL memory devices. Furthermore, we review key technologies for improving the performance of both the CXL memory pool and CXL-based composable computing architecture along with various use cases.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.21 no.11
• /
• pp.2850-2861
• /
• 1996

This paper describes the design of AATM switch LIS of shared buffer type with linked-list architecture to control memory access. The proposed switch LSI consists of the buffer memory, controller and FIFO memory blocks and two special circuits to avoid the cell blocking. One of the special circuit is a new address control scheme with linked-list architecture which maintains the address of buffer memory serially ordered from write address to read address. All of the address is linked as chain is operated like a FIFO. The other is slip-flag register it will be hold the address chain when readaddress missed the reading of data. The circuits control the buffer memory efficiently and reduce the cell loss rate. As a result the designed chip operates at 33ns and occupied on 2.7*2.8mm$^{2}$ using 0.8.mu.m CMOS technology.

• International journal of advanced smart convergence
• /
• v.12 no.1
• /
• pp.70-75
• /
• 2023

With the recent proliferation of memory-intensive workloads such as deep learning, analyzing memory access characteristics for efficient memory management is becoming increasingly important. Since read and write operations in memory access have different characteristics, an efficient memory management policy should take into accountthe characteristics of thesetwo operationsseparately. Although some previous studies have considered the different characteristics of reads and writes, they require a modified hardware architecture supporting read bits and write bits. Unlike previous approaches, we propose a software-based management policy under the existing memory architecture for considering read/write characteristics. The proposed policy logically partitions memory space into the read/write area and the write area by making use of reference bits and dirty bits provided in modern paging systems. Simulation experiments with memory access traces show that our approach performs better than the CLOCK algorithm by 23% on average, and the effect is similar to the previous policy with hardware support.

• Journal of the Korea Society of Computer and Information
• /
• v.25 no.1
• /
• pp.1-11
• /
• 2020

In this paper, we present a technique that executes MPI parallel programs, that are developed on processor-centric computer architecture, more efficiently on memory-centric computer architecture without program modification. The technique we present here improves performance by replacing low-speed data communication over the network of MPI library functions with high-speed data communication using the property called fast large shared memory of memory-centric computer architecture. The technique we present in the paper is implemented in two programs. The first program is a modified MPI library called MC-MPI-LIB that runs MPI parallel programs more efficiently on memory-centric computer architecture preserving the semantics of MPI library functions. The second program is a simulation program called MC-MPI-SIM that simulates the performance of memory-centric computer architecture on processor-centric computer architecture. We developed and tested the programs on distributed systems environment deployed on Docker based virtualization. We analyzed the performance of several MPI parallel programs and showed that we achieved better performance on memory-centric computer architecture. Especially we could see very high performance on the MPI parallel programs with high communication overhead.