WJWooyoung Jo

Wooyoung Jo

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Wooyoung Jo has a total of 28 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • MIDAS: An Energy-Efficient Microscaling Quantization-Based Digital CIM Accelerator With Spatio-Temporal Cyclic Alignment for Generative AI

    by Jiwon Choi, Seongyon Hong, Wooyoung Jo, Wonhoon Park, Sunjoo Whang, Seryeong Kim, Sangjin Kim, Hoi-Jun Yoo

    Digital compute-in-memory (DCIM) and the microscaling (MX) data format have emerged as promising hardware and algorithmic solutions, respectively, for energy-efficient generative AI (GenAI) processing. However, their combination introduces challenges, including redundant computation from mantissa alignment, excessive exponent handling overhead, and accuracy loss from static approximation schemes that fail to accommodate block-wise dynamic bit-significance. In this article, we present MIDAS, an MX-optimized DCIM-based GenAI accelerator with three key features: 1) spatio-temporal cyclic alignment (STCA) with sign-magnitude-based assembled column addition (SACA) enables compact weight mapping, high input reusability, and low-latency computation, achieving $1.72\times $ higher macro energy efficiency; 2) MX-aware minimum exponent post-detection (MEPD) eliminates redundant maximum exponent detection by exploiting the inherent characteristics of MX quantization, reducing exponent handling power, area, and latency by 68.8%, 63.2%, and 22.5%, respectively; and 3) sense amplifier-based dynamic bit (SADB)-significance handling with shared scale-aware channel regrouping (SSCR) dynamically adjusts approximation bit-positions according to each block’s shared scale, improving energy efficiency by 27.3% while minimizing accuracy loss. MIDAS is fabricated in Samsung 28-nm CMOS technology and occupies a 7.56 mm2 die area. MIDAS achieves macro energy and area efficiencies of 96.3 TFLOPS/W and 0.69 TFLOPS/mm2, respectively, under MXFP4. Compared to state-of-the-art DCIMs, MIDAS demonstrates $1.58\times $ higher system energy efficiency and $3.32\times $ higher macro energy efficiency, validating its effectiveness as an MX-optimized AI accelerator for next-generation GenAI applications.

  • SMoLPU: An Energy-Efficient Sparse Mixture-of-Experts-Based Speculative Decoding Language Processing Unit

    by Sangwoo Ha, Jingu Lee, Youngjin Moon, Sunjoo Whang, Wooyoung Jo, Gwangtae Park, Sangjin Kim, Soyeon Um, Junha Ryu, Yurim Jo, et al.

  • C-Transformer: An Energy-Efficient Homogeneous DNN-Transformer/SNN-Transformer Processor for Large Language Models

    by Sangyeob Kim, Sangjin Kim, Wooyoung Jo, Soyeon Kim, Seongyon Hong, Nayeong Lee, Jungwan Lee, Hoi-Jun Yoo

    In this article, we propose a new language model processor named the C-Transformer to address the external memory bottleneck of language models. It consists of three key functional blocks: 1) homogeneous deep-neural-network (DNN)–Transformer/spiking-Transformer core (HDSC) with a hybrid multiplication/accumulation unit (HMAU) to enhance hardware utilization; 2) output spike speculation unit (OSSU) to increase the energy efficiency of spike-domain processing; and 3) implicit weight generation unit (IWGU) with extended sign compression (ESC) to eliminate the external memory bottleneck. The chip is fabricated in Samsung’s 28-nm 1P8M CMOS technology and operates at a supply voltage of 0.7–1.1 V with a maximum frequency of 200 MHz and supports various tasks such as language modeling, translation, and summarization. For models like generative pre-trained transformer 2 (GPT-2), multilingual text-to-text transfer transformer (mT5), text-to-text transfer transformer (T5), and FairSeq MachineTranslation (FSMT), the C-Transformer achieves $0.21{\times }$ – $0.33{\times }$ computation energy and $0.37{\times }$ – $0.41{\times }$ external memory access (EMA) energy compared to the baseline. Our chip demonstrates 13.6% lower energy consumption than the previous state-of-the-art, despite having $2.1{\times }$ more parameters. Moreover, it consumes 63.8% less energy with a similar parameter size. The C-Transformer can complete various language model tasks with <1 s latency, notably FSMT in 0.09 s and GPT-2 in 0.656 s. By combining DNN-Transformer and Spiking-Transformer architectures, the C-Transformer enhances computational energy efficiency, eliminates external memory bottlenecks, and enables language models such as GPT-2 to achieve state-of-the-art performance on mobile devices.


  • Dyamond: Compact and Efficient 1T1C DRAM IMC Accelerator With Bit Column Addition for Memory-Intensive AI

    by Seongyon Hong, Wooyoung Jo, Sangjin Kim, Sangyeob Kim, Soyeon Um, Kyomin Sohn, Hoi-Jun Yoo

    This article proposes Dyamond, a one transistor, one capacitor (1T1C) dynamic random access memory (DRAM) in-memory computing (IMC) accelerator with architecture-to-circuit-level optimizations for high memory density and energy efficiency. The bit column addition (BCA) dataflow introduces output bit-wise accumulation to exploit varying accuracy and energy characteristics across different bit positions. The lower BCA (LBCA) reduces analog-to-digital converter (ADC) operations to enhance energy efficiency with inter-column analog accumulation. The higher BCA (HBCA) improves accuracy through signal enhancement and minimizes energy consumption per ADC readout with signal shift (SS). The design maximizes memory density by dedicating 1T1C cells solely to memory and integrating a compact computation circuit adjacent to the bitline sense amplifier. The memory access power is further reduced with a big-little array structure and a switchable sense amplifier (SWSA), which trades off retention time and energy consumption. Fabricated in 28-nm CMOS, Dyamond integrates 3.54-MB DRAM in a 6.48-mm2 area, achieving 27.2 TOPS/W peak efficiency and outstanding performance in advanced models such as BERT and GPT-2.


  • Scaling-CIM: eDRAM In-Memory-Computing Accelerator With Dynamic-Scaling ADC and Adaptive Analog Operation

    by Sangjin Kim, Soyeon Um, Wooyoung Jo, Jingu Lee, Sangwoo Ha, Zhiyong Li, Hoi-Jun Yoo

    This article presents Scaling-computing-in-memory (CIM), an energy-efficient embedded dynamic random access memory (eDRAM)-based in-memory-computing (IMC) accelerator with a dynamic-scaling readout for signal-to-quantization-noise ratio (SQNR) boosting and analog-to-digital converter (ADC) overhead reduction. It greatly saves the ADC cost by reducing the required number of ADC-bit and ADC operations by codesigning the algorithm and hardware. Scaling-CIM proposes three key features: 1) dynamic scaling ADC (DSA) boosts SQNR of multibit operation even with low-bit ADC; 2) adaptive analog bit-parallel (AABP) accumulation reduces the redundant ADC operation; and 3) layer-wise adaptive bit-truncation (LABT) search further enhances efficiency on benchmarks. The Scaling-CIM is fabricated in 28-nm CMOS technology and occupies a 2.03-mm2 die area with an 800-kb eDRAM cell. It achieves 39.7-TOPS/W (8–9 b) energy efficiency on the RestNet-18 benchmark and $1.96\times $ higher efficiency figure of merit (FoM) than the previous IMC-based accelerator.


  • Top Cited Papers

    DynaPlasia: An eDRAM In-Memory Computing-Based Reconfigurable Spatial Accelerator With Triple-Mode Cell

    by Sangjin Kim, Zhiyong Li, Soyeon Um, Wooyoung Jo, Sangwoo Ha, Juhyoung Lee, Sangyeob Kim, Donghyeon Han, Hoi-Jun Yoo

    This article presents DynaPlasia, a reconfigurable eDRAM-based in-memory computing (IMC) processor with a novel triple-mode cell. It enables higher system-level performance and efficiency in a resource-limited environment. DynaPlasia proposes five key features that can enhance the energy efficiency and area efficiency of the IMC accelerator: 1) dynamic reconfigurable core architecture (DRECA), which dynamically reconfigures the effective IMC macro size according to DNN workloads; 2) the triple-mode cell is reconfigured as IMC PE, unit DAC, and unit cell to optimize system resource utilization adaptively; 3) hierarchical in-memory ADC (Hi-ADC) reduces ADC area overhead with tripe-mode cell and saves power with the hierarchical operation; 4) signed-input signed-weight (SISW) IMC reduces logic switching activity with signed magnitude representation; and 5) leakage-tolerant computing (LTC) to enable high-density cell design. The DynaPlasia is fabricated in 28-nm CMOS technology and occupies a 20.25-mm 2 die area with 9.6-Mb cell capacity. Its measured system peak throughput is 19.5 TOPS (4 b–5 b), and system peak efficiency is 56.0 TOPS/W (4 b–5 b). Even with the ResNet50 @ ImageNet dataset, system efficiency is 29.5 TOPS/W which is 2.5 $\times$ higher than the previous state-of-the-art.


  • A 92 fps and 2.56 mJ/Frame Computing-In-Memory-Based Human Pose Estimation Accelerator With Resource-Efficient Macro for Mobile Devices

    by Beomseok Kwon, Zhiyong Li, Sangjin Kim, Wooyoung Jo, Hoi-Jun Yoo

    Human pose estimation (HPE) is a promising solution for accurately understanding the state and context of human actions in virtual reality (VR). A high frame rate with low-power HPE processing is required for a realistic user interaction experience in battery-limited mobile devices. The proposed HPE accelerator is a computing-in-memory (CIM) based accelerator that computes depth-wise separable convolution (DWSC) of a lightweight HPE network. Three key features contribute to a resource-efficient CIM accelerator: 1) Dual-mode CIM computes DWSC with a reconfigurable homogenous architecture, resulting in 2.68 times higher throughput than previous analog CIMs. 2) Effective layeraware unrolling performs bit-parallel computation on dualmode CIM with fewer ADC operations, achieving 46 times higher throughput than before. 3) Adaptive fused inter-macro balancing improves latency balance in layer fusion execution, leading to a 57.0 % higher frame rate than before. The proposed HPE accelerator is implemented in 28nm CMOS technology. It achieves higher computation resource utilization and operates HPE with a low energy-delay product of 27.6 uJ7s in mobile VR devices.

  • OmniDRL: An Energy-Efficient Deep Reinforcement Learning Processor With Dual-Mode Weight Compression and Sparse Weight Transposer

    by Juhyoung Lee, Sangyeob Kim, Sangjin Kim, Wooyoung Jo, Ji-Hoon Kim, Donghyeon Han, Hoi-Jun Yoo

    In this article, we present an energy-efficient deep reinforcement learning (DRL) processor, OmniDRL, for DRL training on edge devices. Recently, the need for DRL training is growing due to the DRL’s distinct characteristics that can be adapted to each user. However, a massive amount of external and internal memory access limits the implementation of DRL training on resource-constrained platforms. OmniDRL proposes four key features that can reduce external memory access by compressing as much data as possible and can reduce internal memory access by directly processing compressed data. A group-sparse training (GST) enables a high weight compression ratio (CR) for every DRL iteration by selective utilization of weight grouping and weight pruning. A group-sparse training core is proposed to fully take advantage of compressed weight from GST by skipping redundant operations and reusing duplicated data. An exponent-mean-delta encoding additionally compresses the exponent of both weight and feature map for higher CR and low memory power consumption. A world-first on-chip sparse weight transposer enables the DRL training process of compressed weight without off-chip transposer. As a result, OmniDRL is fabricated in a 28-nm CMOS technology and occupies a $3.6\times 3.6$ mm 2 die area. It shows a state-of-the-art peak performance of 4.18 TFLOPS and a peak energy efficiency of 29.3 TFLOPS/W. It achieved 7.42-TFLOPS/W energy efficiency for training robot agent (Mujoco Halfcheetah, TD3), which is 2.4 $\times $ higher than the previous state of the art.


  • TSUNAMI: Triple Sparsity-Aware Ultra Energy-Efficient Neural Network Training Accelerator With Multi-Modal Iterative Pruning

    by Sangyeob Kim, Juhyoung Lee, Sanghoon Kang, Donghyeon Han, Wooyoung Jo, Hoi-Jun Yoo

    This article proposes the TSUNAMI, which supports an energy-efficient deep-neural-network training. The TSUNAMI supports multi-modal iterative pruning to generate zeros in activation and weight. Tile-based dynamic activation pruning unit and weight memory shared pruning unit eliminate additional memory access. Coarse-zero skipping controller skips multiple unnecessary multiply-and-accumulation (MAC) operations at once, and fine-zero skipping controller skips randomly located unnecessary MAC operations. Weight sparsity balancer solves a utilization degradation caused by weight sparsity imbalance, and the workload of each convolution core is allocated by a random channel allocator. The TSUNAMI achieves an energy efficiency of 3.42 TFLOPS/W at 0.78V and 50MHz with floating-point 8-bit activation and weight. Also, it achieves an energy efficiency of 405.96 TFLOPS/W at 90% sparsity condition.