SYSungwoong Yune

Sungwoong Yune

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Sungwoong Yune has a total of 25 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • A 57.3-fps 12.8 TFLOPS/W Text-to-Motion Processor With Inter-Iteration Output Sparsity and Inter-Frame Joint Similarity

    by Jaehoon Heo, Adiwena Putra, Sungwoong Yune, Jieon Yoon, Hangyeol Lee, Ji-Hoon Kim, Joo-Young Kim

    Recently, 3-D human motion generation has become essential in media applications such as film production and augmented reality (AR)/virtual reality (VR) devices, requiring the generation of human joint movements and detailed 3-D meshes for each joint. Traditionally, joint creation required hours or even days, making it impractical for real-time applications. Although the emergence of diffusion models has reduced execution time to within seconds, challenges remain for real-time deployment on edge devices, including high computational cost stemming from denoising iterations and the lengthy processing time of mesh construction based on the skinned multi-person linear (SMPL) model. To address these challenges, we present HuMoniX, a software–hardware (SW–HW) co-designed text-to-motion processor for low-power, real-time 3-D human motion generation. At the software level, we exploit inter-iteration output sparsity in joint creation and inter-frame joint similarity in mesh construction. Computational costs in diffusion models are reduced by skipping redundant operations across iterations in the feed-forward network (FFN) layers, while SMPL is executed only on frames with significant motion, thereby reducing latency. At the hardware level, HuMoniX comprises two heterogeneous engine clusters. The joint creation cluster accelerates matrix multiplication (MMUL) on compacted data through the ConMerge mechanism, effectively managing output sparsity via condensing and merging. The mesh construction cluster focuses on selected frames, employing mixed-precision floating-point (FP) arithmetic to further minimize computational overhead. Fabricated using 14-nm technology, the chip occupies a die area of 12 mm2 and operates at frequencies of 50–600 MHz, consuming 39.18–1382.39 mW at 0.63–0.94 V. It successfully generates human motions from text or music inputs across three benchmarks with minimal accuracy loss. Notably, it achieves a state-of-the-art speed of 57.3 frames/s (fps), making it up to $650.5\times $ faster than comparable GPU systems. Moreover, its energy efficiency reaches 8.9–11.0 TOPS/W for joint creation and 12.3–12.8 TFLOPS/W for mesh construction, outperforming GPUs by $1729\times $ to $3572\times $ and CPUs by 2–6M times. These results establish HuMoniX as a highly efficient and practical solution for resource-constrained edge environments.


  • DPIM: A 2T1C eDRAM Transformer-in-Memory Chip With Sparsity-Aware Quantization and Heterogeneous Dense–Sparse Core

    by Donghyuk Kim, Jae-Young Kim, Hyunjun Cho, Seungjae Yoo, Sukjin Lee, Sungwoong Yune, Sejeong Yang, Hoichang Jeong, Keonhee Park, Ki-Soo Lee, et al.

    Transformer models have revolutionized artificial intelligence (AI) applications across various domains, but their increasing complexity poses significant challenges in terms of computational and memory demands. While processing-in-memory (PIM) paradigms have been adopted to address these limitations, existing PIM-based transformer accelerators still face hurdles such as: 1) focusing solely on optimizing attention layers; 2) lack of sparsity exploitation for transformers; and 3) limited PIM macro capacity and low cell density, which degrades on-chip data reuse and increases external memory access (EMA). This article presents DPIM, a novel 2T1C eDRAM-based transformer-in-memory chip that addresses these challenges through three key innovations: 1) a sparsity-aware quantization (SAQ) scheme that significantly increases bit-slice sparsity in both activation and weight data, achieving ratios of 83.3% and 88.4%, respectively, with minimal accuracy loss; 2) a heterogeneous PIM core capable of efficiently handling both sparse and dense matrix multiplications (MMs); and 3) a high-density 2T1C eDRAM cell with a density of 1.38 Mb/mm2, enabling large-capacity PIM macros. By integrating these features, DPIM achieves improved computational efficiency and reduced EMA with enhanced on-chip data reuse. The DPIM chip, fabricated using 28-nm CMOS technology, achieves a throughput of 3.03–12.12 TOPS and an energy efficiency of 4.84–19.36 TOPS/W, all measured across INT8 and INT4 operations, respectively. It achieves a throughput density of 0.55 TOPS/mm2 with INT8 operation. With a total macro size of 4608 kb, the chip occupies a die area of 20.25 mm2 and operates at frequencies from 50 to 285 MHz with a supply voltage of 0.85–1.0 V. The DPIM successfully executes BERT-Large on the general language understanding evaluation (GLUE) dataset. Its macro density is 1413 kb/mm2, and the resulting density figure-of-merit (FoM) (macro density $\times $ throughput density) is $1.6\times $ – $115.8\times $ higher than previous works, representing a significant advancement in hardware design for efficient transformer processing.