Zongle Huang

Zongle Huang

Auto-generated profile, based on publicly available information

Zongle Huang

Department of Electronic Engineering, Tsinghua University, Beijing, China

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
18 August 2025 in IEEE Journal of Solid-State Circuits
Verified
Journal Article
Closed access
YHYifan He
Yifan He
SFShupei Fan
Shupei Fan
XLXuan Li
Xuan Li
LLLuchang Lei
Luchang Lei
WJWenbin Jia
Wenbin Jia
Yongpan LiuYongpan Liu
Yongpan Liu
Digital circuit is a promising approach to implement computing-in-memory (CIM) architecture for data-intensive applications, such as neural network inference. Previous digital CIM implementations have demonstrated higher throughput and robustness than the analog counterpart, benefiting from technology scaling and full-precision datapaths. However, the traditional digital CIM designs adopt a bulky adder tree as the computing circuit, limiting further improvements in area and energy efficiency. To address the above limitation, this work proposes a novel design methodology that combines the lookup-table (LUT) and high-density embedded dynamic RAM (eDRAM) cells. By storing precomputed weight summations in LUTs, the proposed approach reduces the first two stages of the adder tree. Additionally, the row-level parallelism of the CIM architecture is utilized to amortize the overheads of LUT weight encoding and eDRAM refreshing. The 28-nm prototype demonstrates a peak area efficiency of 16.2 TOPS/mm2 and an energy efficiency of 49.5 TOPS/W, both for 8-bit operations. The proposed eDRAM macro achieves a 2.4 and 0.48 Mb/mm2 density in memory and CIM mode, due to its balanced computing-storage ratio, enabling reconfigurability between computing cores and memory storage.

29 December 2025 in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
Verified
Journal Article
Closed access
WWWenxun Wang
Wenxun Wang
LMLikai Ma
Likai Ma
CTChen Tang
Chen Tang
ZHZongle Huang
Zongle Huang
BCBoju Chen
Boju Chen
Yongpan LiuYongpan Liu
Yongpan Liu
Diffusion models have shown marked advancements in 2D generation and also become focal points in 3D generation via consistent multi-view images generation. However, the computation and memory demands hinder their real-time deployment on mobile and edge devices. Moreover, the reduction of diffusion timesteps leads to a decrease in inter-step similarity, thereby making previous methods ineffective in optimizing computation. The varied layer quantization sensitivity and optimal format in diffusion models also present challenges for traditional quantization methods to achieve efficient memory compression. To address these issues, we first introduce M3 quantization, a mixed-precision, mixed-format, and mixed-granularity quantization framework that allocate optimal precision and format to different data with an unified FP8 computation flow for efficient implementation. Building upon M3 quantization, we present TriM-D, a hardware accelerator designed to optimize computation. It features sparsity-aware dual-branched FP8-MAC units that introduces dual-branch path and efficient approximate adder to fully utilize bit-level sparsity of M3-quantized data. Such FP8-MAC scheme preserves both efficiency and numerical accuracy. Our experiments demonstrate that TriM-D significantly outperforms the state-of-the-art (SOTA) diffusion accelerator Cambricon-D, achieving an average reduction of 56.5% in memory access and 80.6% in computation cost. Additionally, TriM-D provides a 55x, 6.72x improvement in energy-efficiency and 3.69x, 2.67x speedups over NVIDIA A100 and Cambricon-D.

11 November 2025 in IEEE Journal of Solid-State Circuits
Verified
Journal Article
Closed access
CTChen Tang
Chen Tang
ZHZongle Huang
Zongle Huang
WWWenxun Wang
Wenxun Wang
XFXiaoyu Feng
Xiaoyu Feng
YHYifan He
Yifan He
Yongpan LiuYongpan Liu
Yongpan Liu
Diffusion-based models have become increasingly prevalent in content creation applications, significantly shaping the future of entertainment, education, and business marketing. In addition to the primary denoising network, these models often incorporate transformer-based modules for customized generation and post-processing stages to ensure high-quality media outputs. However, the diversity of model architectures, coupled with their substantial computational and memory demands, poses a major challenge for efficient inference on personal devices. Facing these challenges, this work proposes CCE, a content creation engine that provides unified acceleration for diffusion-based applications. CCE decomposes activation into outliers and dense features based on the shared task feature, enabling block-wise outlier processing and efficient differential computation for video generation and upscaling. Furthermore, CCE introduces an LUT-based tensor core featuring a bipolar lookup mechanism, which supports mixed-representation computing of different data formats and precision. In addition, a semantic-driven instruction generation unit is developed to compress activation, further enhancing the system’s efficiency adaptively. With these innovations, CCE achieves up to an 84% reduction in memory requirement and $7{\times }$ acceleration. Fabricated using 28-nm CMOS technology, CCE demonstrates $3.58{\times }$ , 28%, and $2{\times }$ higher average energy efficiency compared to prior diffusion, transformer, and super resolution (SR) accelerators, respectively.

7 November 2026
Verified
Conference Paper
Open access
ZHZongle Huang
Zongle Huang
SFShupei Fan
Shupei Fan
XLXinyuan Lin
Xinyuan Lin
CTChen Tang
Chen Tang
WCWenya Chen
Wenya Chen
Yongpan LiuYongpan Liu
Yongpan Liu
Training LLM requires hybrid parallelism across thousands of GPUs, yet tensor parallelism (TP)—partitioning large weight tensors—remains limited to small degrees in practice due to hardware and software scalability constraints. This limited TP degree forces deeper pipeline parallelism or larger data parallelism, each incurring additional inefficiencies. It also increases memory bandwidth demand, placing an unnecessary burden on already capacity-constrained training systems. Driven by these limitations, we advocate scalable TP, where the TP degree scales with the model hidden size. We realize this with Argus, a communication- and memory-efficient system co-designed across hardware and software. On the hardware front, Argus leverages a switchless chiplet architecture for low link latency and high scalability. We further introduce a split-memory design that consistently prevents memory access from becoming a bottleneck at scale under physical and cost constraints. On the software front, we develop a TP solution with low communication complexity, and adopt an interleaved transfer pattern tailored for connectivity-limited switchless networks. We further remove the TP-degree constraint imposed by attention head count by eliminating partial-sum transfer overhead. Collectively, these optimizations enable coordinated scaling: computation, communication, and memory overheads remain asymptotically constant as the TP degree increases. This property is validated by both theoretical analysis and experiments, with Argus delivering 3.10×/2.02× speedup/energy saving over typical TP baselines. Scalable TP further improves hybrid-parallel training efficiency by better balancing different forms of parallelism, yielding an estimated 65% MFU for Llama3.1-405B.

16 April 2025 in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
Verified
Journal Article
Closed access
LHLiu He
Liu He
YWYujin Wang
Yujin Wang
ZHZongle Huang
Zongle Huang
SFShupei Fan
Shupei Fan
CTChen Tang
Chen Tang
HJHongyang Jia
Hongyang Jia
Transformer-based neural networks (NNs) have been dominant in modern AI due to their superior accuracy and generalization over traditional models. However, the quadratic complexity and dataflow challenges of self-attention hinder their efficient deployment on edge devices, especially for latency-sensitive, single-batch inference. Limited data parallelism necessitates both pipeline and tensor parallelism. Specifically, sequence parallelism offers strong scaling potential in domain-specific accelerators (DSAs), but is constrained by the temporal dependencies involved in the max-finding operations of softmax. This work presents SASDenSebLE, a compact, scalable vision Transformer (ViT) inference architecture co-designed with a max-finding-free softmax approximation across micro-architecture, architecture, and algorithm levels. A dense-sparse-hybrid datapath transforms large-value softmax operations into linear projections, enabling layer-fusion. A hierarchical system further improves sequence parallelism without excessive on-chip buffering or off-chip traffic. Experiments show up to $2.83\times $ , $28.02\times $ , and $1.40\times $ speedup over baseline DSAs and a state-of-the-art ViT accelerator design, respectively, with negligible accuracy loss. Demonstration on the camera encoder of a recent bird’s-eye-view model shows up to $10.82\times $ speedup over a typical edge GPU.

27 May 2026 in IEEE Transactions on Circuits and Systems I: Regular Papers
Verified
Journal Article
Closed access
CTChen Tang
Chen Tang
XLXinyuan Lin
Xinyuan Lin
ZHZongle Huang
Zongle Huang
WSWenyu Sun
Wenyu Sun
HJHongyang Jia
Hongyang Jia
Yongpan LiuYongpan Liu
Yongpan Liu
Transformer-based neural networks are pivotal to emerging vision and natural language processing tasks. Despite their success, accelerating these models remains challenging due to their immense computational and memory requirements. Traditional methods, such as sparse attention computing, alleviate some of these burdens but fall short in addressing the complexity of diverse operators and the significant storage overhead caused by large model parameters and low data reuse rates. In this work, we propose BVDT, an energy-efficient Transformer accelerator that exploits inherent matrix redundancy to overcome these limitations through four key innovations. First, BVDT introduces a novel basis-vector decomposition technique that compresses model size by over $25\times $ , supported by unified memory management. Then, it incorporates a dual array core architecture, powered by LUT-assisted computing, featuring an optimized lookup code fetch and LUT programming mechanism. In addition, BVDT employs a block-wise post-fusion mapping method that seamlessly links the compression pattern and computing architecture. Finally, BVDT utilizes a private-to-private highway data-path for multi-core systems, enabling a LUT stationary dataflow that significantly reduces the memory access overhead. Consequently, BVDT provides a task-agnostic optimization across diverse Transformer operators under various workloads. Fabricated in 28-nm CMOS technology, the accelerator achieves an exceptional area efficiency of $4.35~TOPS/mm^{2}$ , a $5.1\times $ improvement over prior state-of-the-art solutions. Moreover, BVDT also achieves $1.4\times $ to $3.6\times $ regarding energy efficiency, offering a promising solution for Transformer-based applications.