DWDanghui Wang

Danghui Wang

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Danghui Wang has a total of 24 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Maestro: Learning-Based Reuse of Activated Experts to Mitigate Long-Tail Latency under Heavy MoE Workloads in Cloud

    by Jihe Wang, Yulu Liu, Yuanhua Xiao, Linying Wu, Zhiyu Sun, Jianfeng An, Danghui Wang

    Cloud LLM services increasingly adopt Mixture-of-Experts (MoE) models for cost-effective scaling, yet under heavy multi-task workloads their serving performance is often bottlenecked by severe tail latency. The execution semantics of MoE inference are poorly aligned with cloud-native serving abstractions, causing spatial and temporal mismatch in expert placement and demand, and consequently frequent expert loading. Expert demand, however, exhibits structured locality where correlated requests and decoding steps often induce overlapping expert working sets, creating opportunities for activated-expert reuse. Exploiting such reuse is non-trivial because request routing, expert reuse, and resource allocation are tightly coupled over time, forming a sequential decision problem whose coupled dimensions existing systems address only in isolation. This paper presents Maestro, a learning-based framework for activated-expert reuse in cloud-native MoE inference. Maestro models tasks, experts, instances, and execution stages as a heterogeneous graph with relational message passing to capture structured system state, and combines a lightweight Neural Optimistic Value Policy (NOVP) with Pareto-guided action ranking (AdaFilter) for efficient online decision making. It further introduces a prefill-driven expert prefetch mechanism that overlaps loading with computation, where the per-window transfer budget is jointly determined by the RL controller to balance reuse coverage against memory pressure. Evaluated on Mixtral-8 × 7B, Phi-3.5-MoE, and Qwen1.5-MoE, Maestro reduces P50 and P99 latency by up to 34.02% and 28.94%, respectively, and achieves up to 1.51 × the throughput of vLLM. Against the SOTA learning-based methods, it reduces P99 and TTFT by 8.58%–17.47% and 6.22%–17.43%, respectively, while improving throughput by 12.09%–22.70%.

  • Open Access

    SGAP: A Scatter-Gather Processing-in-Memory Architecture with Fixed-Point Reconfiguration for Energy-Efficient DNNs

    by Jihe Wang, Jiawang Cao, Feilong Qiaoyuan, Zhaoqing Wang, Jianfeng An, Danghui Wang

    Leveraging the robustness of deep neural networks (DNNs), fixed-point mixed-precision models have effectively reduced hardware costs in edge in-memory accelerators. However, existing studies on multipliers are poorly suited to mixed-precision models. On the one hand, most of them employ fixed bit-widths or offer only limited configurations, lacking the fine-grained reconfigurability required by mixed-precision models. On the other hand, layers with smaller retained bit-widths can tolerate larger multiplication errors to achieve lower hardware costs, as the inaccurate bits are discarded. However, existing approximate multipliers cannot achieve low-cost dynamic error control, resulting in this optimization potential being wasted. To address these issues, this study proposes a fixed-point approximate multiplier with fine-grained reconfigurability. By breaking full multiplication into small 2-bit/3-bit operations and discarding non-critical ones selectively, the proposed multiplier ensures high hardware utilization across varying bit-widths while enabling low-cost fine-grained error control. Building on this multiplier, a DDR5-oriented near-memory accelerator architecture is proposed, incorporating extended dynamic random-access memory (DRAM) commands and semi-SIMD scheduling schemes. Compared to state-of-the-art processing-in-memory (PIM) architectures, the proposed architecture achieves a reduction of 40.8% in area, 16.3% in power consumption, and 6.5% in latency, respectively.

  • Strain-gradient-driven synthetic antiferromagnetic skyrmion neuron with velocity-affine LIF dynamics for neuromorphic computing

    by Zijian Feng, Ru Han, Danghui Wang, Jihe Wang

    This work presents a strain-gradient-driven synthetic antiferromagnetic (SAF) skyrmion neuron that implements leaky-integrate-fire (LIF) functionality through voltage-controlled magnetoelastic actuation. The proposed device comprises a SAF nanotrack elastically coupled to a piezoelectric PZT-5H thin film, where the application of small voltages establishes programmable compressive or tensile strain gradients to drive skyrmion motion. Micromagnetic simulations demonstrate that the SAF configuration effectively suppresses the skyrmion Hall effect, ensuring rectilinear propagation and enabling a precise linear mapping between strain-gradient amplitude and skyrmion velocity. Building on these drift characteristics, a velocity-affine LIF model is formulated, in which integration (K int ) and leakage (V leak ) coefficients are quantitatively extracted from displacement-time trajectories and shown to scale linearly with the applied gradient. The derived neuron model is integrated into the SpikingJelly framework and evaluated within a 784-128-10 spiking neural network trained on the MNIST dataset. Network-level simulations reveal that the skyrmion-based neuron achieves training and inference accuracies exceeding 97%, matching the performance of standard LIF models while exhibiting reduced trial-to-trial variance. To assess hardware-level programmability and design robustness, a two-dimensional parameter sweep across the (K int , V leak ) space is performed. The resulting analysis identifies a broad operational region where inference accuracy remains robust, coinciding with a highly energy-efficient regime of approximately 0.5-1.0 pJ per spike. These results establish a skyrmion neuron architecture that is physically interpretable, energy-efficient, and tunably programmable, bridging nanoscale strain engineering with reliable system-level neuromorphic computation.


  • A Transverse-Read-Assisted Fast Valid-Bits Collection in Stochastic Computing MACs for Energy-Efficient in-RTM DNNs

    by Jihe Wang, Zhiying Zhang, Xingwu Dong, Danghui Wang

    It looks very attractive to coordinate racetrack-memory (RM) and stochastic-computing (SC) jointly to build an ultra-low power neuron-architecture. However, the above combination has always been questioned in a fatal weakness that the heavy valid-bits collection of Racetrack Memory-Magnetic Tunnel Junctions(RM-MTJ), a.k.a. accumulative parallel counters (APCs), cannot physically match the requirement for energy-efficient in-memory DNNs. Fortunately, a recently developed Transverse-Read (TR) provides a lightweight collection of valid-bits by detecting domain-wall resistance between a couple of MTJs on a single nanowire. In this work, we first propose a neuron-architecture that utilizes parallel TRs to build an ultra-fast valid-bits collection scheme specifically targeted at Multiply-Accumulate (MAC) units for in-RTM DNNs, where the multiplication operation inherently involves two operands. To solve the huge storage for full stochastic sequences caused by the limited TR banks, a hybrid coding, pseudo-fractal compression, is designed to generate stochastic sequences by segments. To overcome the misalignment by the parallel early-termination, an asynchronous schedule of TR is further designed to regularize the vectorization, in which the valid-bits from different lanes are merged in multiple RM-stacks for vector-level valid-bits collection. However, an inherent defect of TR, i.e., neighbor parts cannot be accessed simultaneously, could limit the throughput of the parallel vector multiplication, therefore, an interleaving data placement is used for full utilization of the memory bus among different vectors. The experimental results demonstrate that the SC-MAC architecture with TR achieves $2.88\boldsymbol{\times}$–$4.40\boldsymbol{\times}$ speedup over CORUSCANT (the state-of-the-art process-in-RM architecture), while simultaneously reducing energy consumption by $1.26\boldsymbol{\times}$–$1.42\boldsymbol{\times}$.

  • Accelerating Point Cloud Sampling by Parallel Structure Deconstruction

    by Hengzhe Chi, Jihe Wang, Jinzhe Zhang, Danghui Wang

    Point cloud processing is fundamental to applications such as autonomous driving, robotic navigation, and 3D reconstruction. Sampling is a crucial process in point cloud processing, but current sampling algorithms perform poorly when deployed on edge devices due to limited computational resources. This stems from three main reasons. First, the robustness of sampling is weak; even after denoising, residual noise remains, and existing edge sampling methods struggle to ignore noisy sampling nodes. Second, there’s a lack of structural sampling, popular algorithms like FPS tend to cover global information without considering special structures for focused sampling. Third, system efficiency is poor, traditional Farthest Point Sampling (FPS) and k-Nearest Neighbor Search (kNN) have terrible time and space complexities of $O(n^{2})$ with low parallelism, limiting system throughput, especially for large-scale point cloud processing. To bridge this gap, we propose a novel point cloud sampling algorithm based on parallel graph deconstruction named DPS (Deconstruction-based Point Cloud Sampling) and its multi-scale version DPS_MS with better denoising effect. First, during graph construction, we employ multi-scale graphing and add a preprocessing stage to filter out some noise nodes, reducing the noise rate. Second, in the sampling stage, we categorize important structures into edge nodes and dense nodes based on the reverse neighbor count of graph nodes, assigning them higher sampling weights to supplement structural information. Finally, we use a highly parallel locality-sensitive hashing algorithm to accelerate neighbor search and reduce memory consumption, we achieved data-level parallelization through the comprehensive parallelization of our algorithm. Through rigorous qualitative and quantitative validation on classification and segmentation tasks, we demonstrate that DPS, while maintaining accuracy, increases point cloud sampling speed by 22.08 times compared to the FPS algorithm, the implementation achieved a 75.1-fold improvement in system throughput and a maximum parallel speedup of 10.32x, and improves the Accuracy/FLOPs (M) ratio to 8.68.


  • Open Access

    SC-GNN: A Communication-Efficient Semantic Compression for Distributed Training of GNNs

    by Jihe Wang, Ying Wu, Danghui Wang

    Training big graph neural networks (GNNs) in distributed systems is quite time-consuming mainly because of the ubiquitous aggregate operations that involve a large amount of cross-partition communication for collecting embeddings/gradients during the forward and backward propagation. To reduce the volume of the communication, some recent approaches focused on decaying each of connections via sampling, quantifying, or delaying until satisfactory trade-off are obtained between volume and accuracy. However, when applied to popular GNNs, those approaches are found to be bounded by a common volume/accuracy Pareto frontier which shows that the decaying for individual connection cannot further accelerate the aggregate of training. In this work, SC-GNN, a semantic compression of the cross-partition communication, is proposed to concentrate a group of connections as a high-level semantics and transmit to a target partition. Since carrying the overall intent of a group, the semantics can keep transferring the interactions, i.e., embeddings/gradients, between a pair of remote partitions until GNN models converge. In addition, a connection-pattern based differential optimization is proposed to further prune those weak connections, while guaranteeing the training accuracy. The results show that, for multi-field datasets, the compression rate of SC-GNN is 40.8 × higher than SOTA methods and the epoch time is reduced to 31.8% on average.