BNByeong-Gyu Nam

Byeong-Gyu Nam

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Byeong-Gyu Nam has a total of 54 co-authors. Here are the top co-authors by number of shared publications:

  • Hoi-Jun Yoo

    Department of Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Republic of Korea

    20 shared publications
  • Seongrim Choi

    Department of Computer Science and Engineering Chungnam National University Daejeon Korea

    9 shared publications
  • Gyeonghoon Kim

    Department of Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Republic of Korea

    5 shared publications
  • Hyejung Kim

    KAIST, Daejeon

    5 shared publications
  • Jeabin Lee

    Korea Adv. Inst. of Sci. & Technol., Daejeon

    4 shared publications
  • Suhwan Cho

    Department of Computer Science and Engineering Chungnam National University Daejeon Korea

    4 shared publications

Latest Publications

  • Open Access

    FSSA: A Fully Serial and Mixed-Precision Systolic Array for Vision Transformer Acceleration on Edge

    by Iraj Moghaddasi, Byeong-Gyu Nam

    Transformers have achieved remarkable success, often surpassing human performance in specific tasks. However, this noteworthy capability comes at the cost of increasing complexity, posing significant challenges when deploying on resource-constrained edge devices. Concerning this, previous efforts have primarily focused on enhancing computational efficiency via compression. Leading-edge studies often utilize approximate computing to improve efficiency with minimal loss of accuracy. Meanwhile, mixed-precision accelerators enhance efficiency while preserving accuracy, albeit with extra overhead. Conversely, bit-serial processing presents a compelling alternative to improve efficiency without sacrificing accuracy or increasing overhead. This article introduces the Fully Serial Systolic Array (FSSA) for Transformer acceleration on edge devices. The proposed architecture integrates bit-serial processing into the systolic array architecture to optimize computational efficiency. FSSA leverages a streamlined design, bit-precision adjustment, and bit-sparsity exploitation to accelerate and enhance efficiency. The proposed architecture is described at the RTL level and synthesized using the 28nm technology library, enabling comprehensive evaluation. Furthermore, the functionality of the FSSA was validated through an FPGA-based implementation. Cycle-accurate simulations of Transformers for image classification reveal that a 16×128 FSSA achieves impressive efficiency gains of 68% and 83% in terms of PDP and EDP, respectively. Additionally, FSSA provides further speed-ups of 31% through bit-precision adjustment and 35% through bit-sparsity exploitation for the same workloads without decreasing accuracy. These results highlight the significance of FSSA, outperforming the bit-parallel baseline in computational efficiency without compromising accuracy.


  • Open Access

    RNPE: An MSDF and Redundant Number System-Based DNN Accelerator Engine

    by Iraj Moghaddasi, Ghassem Jaberipur, Danial Javaheri, Byeong-Gyu Nam

    Deep neural network (DNN) is becoming pervasive in today’s applications with intelligent autonomy. Nonetheless, the ever-increasing complexity of modern DNN models caused several challenges on edge devices, struggling to support the intensive computing demands. Although several hardware accelerators have been introduced to improve performance and computation efficiency, challenges still exist particularly in mission-critical applications, e.g., automotive and healthcare. Compression and approximation have been utilized in this respect, albeit with the probability of accuracy loss. Meanwhile, serial accelerators increase computation efficiency via dynamic precision adaptation and computation pruning but at the expense of increasing response time. This paper proposes the Redundant number system-based Neural Processing Engine (RNPE) with the Most Significant Digit First (MSDF) input and output streams. RNPE reduces the response time while improving computation efficiency compared to traditional bit-parallel and bit-serial processing engines. The proposed architecture has been described in RTL and synthesized in 28 nm CMOS technology for evaluation. Cycle-accurate simulations over the DNN models of image classification demonstrated a single unit of RNPE significantly reduces the response time by up to 97% with no accuracy loss compared to the baseline; however, an additional 25% area overhead is imposed. Furthermore, RNPE improves the average power-delay and energy-delay products by 14% and 53%, respectively. Eventually, RNPE exceeds the state-of-the-art by 23% on average in pruning ineffectual computations on the MSDF output stream.


  • Open Access

    Enhancing Computation-Efficiency of Deep Neural Network Processing on Edge Devices through Serial/Parallel Systolic Computing

    by Iraj Moghaddasi, Byeong-Gyu Nam

    In recent years, deep neural networks (DNNs) have addressed new applications with intelligent autonomy, often achieving higher accuracy than human experts. This capability comes at the expense of the ever-increasing complexity of emerging DNNs, causing enormous challenges while deploying on resource-limited edge devices. Improving the efficiency of DNN hardware accelerators by compression has been explored previously. Existing state-of-the-art studies applied approximate computing to enhance energy efficiency even at the expense of a little accuracy loss. In contrast, bit-serial processing has been used for improving the computational efficiency of neural processing without accuracy loss, exploiting a simple design, dynamic precision adjustment, and computation pruning. This research presents Serial/Parallel Systolic Array (SPSA) and Octet Serial/Parallel Systolic Array (OSPSA) processing elements for edge DNN acceleration, which exploit bit-serial processing on systolic array architecture for improving computational efficiency. For evaluation, all designs were described at the RTL level and synthesized in 28 nm technology. Post-synthesis cycle-accurate simulations of image classification over DNNs illustrated that, on average, a sample 16 × 16 systolic array indicated remarkable improvements of 17.6% and 50.6% in energy efficiency compared to the baseline, with no loss of accuracy.


  • Open Access

    MixedNet: Network Design Strategies for Cost-Effective Quantized CNNs

    by Dong-Jin Chang, Byeong-Gyu Nam, Seung-Tak Ryu

    This paper proposes design strategies for a low-cost quantized neural network. To prevent the classification accuracy from being degraded by quantization, a structure-design strategy that utilizes a large number of channels rather than deep layers is proposed. In addition, a squeeze-and-excitation (SE) layer is adopted to enhance the performance of the quantized network. Through a quantitative analysis and simulations of the quantized key convolution layers of ResNet and MobileNets, a low-cost layer-design strategy for use when building a neural network is proposed. With this strategy, a low-cost network referred to as a MixedNet is constructed. A 4-bit quantized MixedNet example achieves an on-chip memory size reduction of 60% and fewer memory access by 53% with negligible classification accuracy degradation in comparison with conventional networks while also showing classification accuracy rates of approximately 73% for Cifar-100 and 93% for Cifar-10.

  • Open Access

    An area‐efficient memory‐based multiplier powering eight parallel multiplications for convolutional neural network processors

    by Seongrim Choi, Suhwan Cho, Byeong‐Gyu Nam

    Convolutional neural network (CNN) is widely used for various deep learning applications because of its best‐in‐class classification performance. However, CNN needs several multiply‐accumulate (MAC) operations to realize human‐level cognition capabilities. In this regard, an area‐efficient multiplier is essential to integrate a large number of MAC units in a CNN processor. In this letter, we present an area‐efficient memory‐based multiplier targeting CNN processing. The proposed architecture adopts a 32‐port memory shared across eight multiplications. Simulation results show that area is reduced by 18.4% compared with the state‐of‐the‐art memory‐based multiplier.


  • Compact Mixed-Signal Convolutional Neural Network Using a Single Modular Neuron

    by Dong-Jin Chang, Byeong-Gyu Nam, Seung-Tak Ryu

    This paper demonstrates a compact mixed-signal (MS) convolutional neural network (CNN) design procedure by proposing a MS modular neuron unit that alleviates analog circuit related design issues such as noise. Through the first step of the proposed procedure, we design a CNN in software with a minimized number of channels for each layer while satisfying the network performance to the target, which creates low representational and computational cost. Then, the network is reconstructed and retrained with a single modular neuron that is recursively utilized for the entire network for the maximum hardware efficiency with a fixed number of parameters that consider signal attenuation. For the last step of the proposed procedure, the parameters of the networks are quantized to an implementable level of MS neurons. We designed the networks for MNIST and Cifar-10 and achieved compact CNNs with a single MS neuron with 97% accuracy for MNIST and 85% accuracy for Cifar-10 whose representational cost and computational cost are reduced to least two times smaller than prior works. The estimated energy per classification of the hardware network for Cifar-10 with a single MS neuron, designed with optimum noise and matching requirements, is 0.5μ J, which is five times smaller than its digital counterpart.


  • A self-powered always-on vision-based wake-up detector for wearable gesture user interfaces

    by Suhwan Cho, Seongrim Choi, Junsik Woo, Ara Kim, Byeong-Gyu Nam

    Hand gesture recognition is one of the secure natural user interface (NUI) mechanisms on wearable devices since it does not reveal user's intention in public domain unlike the speech recognition. However, its energy dissipation is very demanding as it requires compute-intensive machine vision processing. Recently, wake-up detectors have been proposed to improve the energy-efficiency of always-on sensing nature of the NUI systems by switching off the main functional blocks while just keeping the wake-up detector alive during idle time. However, vision-based wake-up detectors still require power-consuming vision processing so we propose a self-powered vision wake-up detector to alleviate burdens on limited battery and thus facilitate always-on wake-up detection for the wearable gesture UIs. Our work has four key features to realize the self-powered wake-up detection; 1) imaging-harvesting dual-mode CMOS image sensor (CIS) with 0.6V 3T pixels, 2) subthreshold SRAM with disturb-free 0.3V 10T bitcells, 3) hand detection engine with modified Haar-like filters invariant to skin-colors, and 4) on-die switched capacitor DC-DC converter for lightweight system design. Thanks to the features combined, this work achieves selfpowered operation of always-on vision wake-up detection.


  • A low-power real-time hidden Markov model accelerator for gesture user interface on wearable devices

    by Seongrim Choi, Jaemin Hwang, Suhwan Cho, Ara Kim, Byeong-Gyu Nam

    A low-power and real-time hidden Markov model (HMM) accelerator is proposed for gesture user interfaces on wearable smart devices. HMM algorithm is widely used for sequence recognitions such as speech recognition and gesture recognition due to its best-in-class recognition accuracy. However, the HMM algorithm incorporates high computational complexity and requires massive memory bandwidth for sequence matches. There have been studies on hardware acceleration of the HMM algorithm to resolve these issues, but they were focused on the speech recognition and did not incorporate the motion orientation capability required for the gesture recognition case. In this paper, we propose an HMM accelerator incorporating the motion orientation block for gesture recognitions on wearable devices. Binary search is exploited in the motion orientation to avoid the division and arctangent associated with the orientation and reduce its arithmetic complexity. In addition, gesture models are clustered in the gesture database to save the memory bandwidth by reducing memory transactions. Moreover, logarithmic arithmetic is used in Viterbi decoding in the HMM for more reduction in its complexity. Thanks to these schemes, this work achieves 25.6% power reduction compared with a plain hardware implementation of the gesture recognizing HMM.