arXiv subjects

Guoyu Li

Publications and source records attributed to Guoyu Li.

13 recordsLinked to original sources

arXiv2026

AcceptMoE: Commitment-Weighted Self-Sizing Verifier Expert Sets for Efficient MoE Speculative Decoding

Speculative decoding verifies a tree of draft tokens in one target-model forward pass. For a mixture-of-experts (MoE) target, however, parallel verification can activate the union of the experts selected by all tree nodes, even though only a small subset of those nodes reaches the accepted output. Token count, activated-expert union size, and expert-weight traffic are therefore distinct cost measures: reducing the token workload need not shrink the expert union proportionally, and under offloading, transfer traffic also depends on cache residency. We introduce AcceptMoE, a verifier-side expert selector that combines target-router scores with offline-estimated commitment probabilities and automatically adjusts the number of eligible experts for each verification block, eliminating the need for a user-specified expert budget. Under offloading, AcceptMoE conditions expert eligibility on cache residency instead of predicting natural routes and prefetching the corresponding expert weights. Although constraining target-expert eligibility changes the model distribution, across 12 model-task pairs spanning three MoE targets and four benchmarks, AcceptMoE's mean accuracy is 0.27 percentage points lower than that of EAGLE-3 speculative decoding with natural routing. Served with SGLang at batch size one, it reaches 1.290 times the throughput of this baseline with all expert weights in GPU memory, and 2.06 times under physical expert offloading, while reducing host-to-device traffic by 73.6 percent to 77.1 percent.

cs.LG↗

arXiv2026

TileSight: A First-Principles Tile-Centric Analytical GPU Performance Model from Cores to Clusters

Recent GPU programming frameworks such as Triton, TileLang, and CUDA Tile adopt tiles as first-class primitives, making tile-centric programming the prevailing approach for high-performance GPU kernels. Performance-analysis tooling has not followed: programmers still rely on coarse roofline bounds, opaque ML predictors, or post-hoc profilers to understand kernel execution. This gap is acute for modern AI workloads, where kernel fusion and distributed inference depend on tensor cores, CUDA cores, cache hierarchies, memory pipelines, and inter-GPU networks. We present TileSight, a tile-centric performance-modeling tool that elevates the tile from a programming primitive to an analysis primitive. Within a GPU core, TileSight models compute-memory pipeline overlap; across cores, it models the cache hierarchy; across GPUs, it models inter-node communication. All layers share the tile abstraction: the intra-tile layer expresses work as a resource vector spanning network, memory, and compute pipelines; the inter-tile layer schedules dependent and ordered actions to expose legal overlap and infers multi-level cache hit rates from tile reuse distance; and the cross-device layer maps remote tensor accesses to placements and routes them through an alpha-beta stage cost. On A100, H200, B200, and B6000, TileSight predicts single-GPU kernel latency with 12.35% pooled mean absolute percentage error (MAPE), outperforming state-of-the-art baselines and transferring better across architectures. Its L2 cache-hit-rate predictions are within roughly one percentage point of measurements on every GPU. At up to 32 GPUs, TileSight achieves 16.18% weighted MAPE (wMAPE) on fused distributed kernels and 13.52% wMAPE on end-to-end vLLM serving. In optimization, TileSight selects tile configurations competitive with strong vendor and expert baselines. TileSight will be open-sourced upon publication.

cs.DC↗

arXiv2026

SPAC: Automating FPGA-based Network Switches with Protocol Adaptive Customization

With network requirements diverging across emerging applications, latency-critical services demand minimal logic delay, while hyperscale training and collectives require sustained line-rate throughput for synchronized bulk transfers. This divergence creates an urgent need for custom network switches tailored to specialized protocols and application-specific traffic patterns. This paper presents SPAC (Switch and Protocol Adaptive Customization), a novel approach that automates the generation of FPGA-based network switches co-optimized for custom protocols and application-specific traffic patterns. SPAC introduces a unified workflow with a domain-specific language (DSL) for protocol-architecture co-design, a library of modular HLS-based adaptive switch components, and a trace-aware Design Space Exploration (DSE) engine. By providing a multi-fidelity simulation stack, SPAC enables rapid identification of Pareto-optimal designs prior to deployment. We demonstrate the efficacy of the domain-specific adaptation of SPAC across a spectrum of real-world scenarios, spanning from latency-sensitive sensor and HFT networks to hyperscale datacenter fabrics. Experimental results show that by tailoring the micro-architecture and protocol to the specific workload, SPAC-generated designs reduce LUT and BRAM usage by 55% and 53%, respectively. Compared to fixed-architecture counterparts, SPAC delivers latency reductions ranging from 7.8% to 38.4% across various tasks while maintaining adequate resource consumption and packet drop rate.

cs.NI↗

arXiv2026

DeepStack: Facilitating Co-Design Exploration of 3D DRAM-Stacked Accelerators for Distributed LLM Inference

Advances in hybrid bonding and packaging have driven growing interest in 3D DRAM-stacked AI accelerators. As large language models (LLMs) scale to hundreds of billions or trillions of parameters, distributed inference across multiple 3D chips has become essential for AI serving. This trend makes cross-stack co-design critical because system-level parallelization and scheduling choices are tightly coupled with hardware characteristics such as memory organization, interconnects, and thermal constraints. We present DeepStack, an accurate performance model and efficient design space exploration (DSE) framework for distributed 3D-stacked LLM inference. At the hardware level, DeepStack captures transaction-aware memory bandwidth, bank activation constraints, buffering limitations, and thermal and power behavior. At the system level, it incorporates comprehensive parallelization strategies and execution scheduling. Through a dual-stage network abstraction and tile-level compute-communication overlap modeling, DeepStack achieves up to 100,000x faster evaluation than state-of-the-art simulators at comparable accuracy. We cross-validate DeepStack against our in-house 3D designs, an NS-3 backend with 2.12% error, and vLLM serving on eight B200 GPUs with 12.92% error. Combined with hierarchical search, DeepStack efficiently explores about 2.5 x 10^14 design points spanning the number of stacked DRAM layers, DRAM vertical connectivity, interconnects, compute-memory allocation, and distributed scheduling under thermal and area constraints. A search-space ablation shows that restricted DSE baselines can miss up to 9.5x modeled throughput. Beyond modeling and DSE, DeepStack derives design implications for distributed 3D AI systems and guides performance optimization across the stack. Source code and artifacts are available at https://github.com/tile-ai/DeepStack/tree/ae.

cs.AR↗

arXiv2025

DPUV4E: High-Throughput DPU Architecture Design for CNN on Versal ACAP

Convolutional Neural Networks (CNNs) remain prevalent in computer vision applications, and FPGAs, known for their flexibility and energy efficiency, have become essential components in heterogeneous acceleration systems. However, traditional FPGAs face challenges in balancing performance and versatility due to limited on-chip resources. AMD's Versal ACAP architecture, tailored for AI applications, incorporates AI Engines (AIEs) to deliver high computational power. Nevertheless, the platform suffers from insufficient memory bandwidth, hindering the full utilization of the AIEs' theoretical performance. In this paper, we present DPUV4E for the Versal architecture, providing configurations ranging from 2PE ($32.6$ TOPS) to 8PE ($131.0$ TOPS). We design two computation units, Conv PE and DWC PE, to support different computational patterns. Each computation unit's data flow efficiently utilizes the data reuse opportunities to mitigate bandwidth bottlenecks. Additionally, we extend the functionality of each PE to utilize AIEs for non-convolutional operations, reducing resource overhead. Experiments on over 50 models show that compared to previous designs, our design provides $8.6\times$ the TOPS/W of traditional FPGA-based DPU designs, while reducing DSP usage by $95.8\%$, LUT usage by $44.7\%$, and latency to $68.5\%$ under single-batch conditions. For end-to-end inference, our design improving throughput by up to $2.2\times$ for depth-wise convolution models and up to $1.3\times$ for standard models.

cs.AR↗

arXiv2025

LUT-DLA: Lookup Table as Efficient Extreme Low-Bit Deep Learning Accelerator

The emergence of neural network capabilities invariably leads to a significant surge in computational demands due to expanding model sizes and increased computational complexity. To reduce model size and lower inference costs, recent research has focused on simplifying models and designing hardware accelerators using low-bit quantization. However, due to numerical representation limits, scalar quantization cannot reduce bit width lower than 1-bit, diminishing its benefits. To break through these limitations, we introduce LUT-DLA, a Look-Up Table (LUT) Deep Learning Accelerator Framework that utilizes vector quantization to convert neural network models into LUTs, achieving extreme low-bit quantization. The LUT-DLA framework facilitates efficient and cost-effective hardware accelerator designs and supports the LUTBoost algorithm, which helps to transform various DNN models into LUT-based models via multistage training, drastically cutting both computational and hardware overhead. Additionally, through co-design space exploration, LUT-DLA assesses the impact of various model and hardware parameters to fine-tune hardware configurations for different application scenarios, optimizing performance and efficiency. Our comprehensive experiments show that LUT-DLA achieves improvements in power efficiency and area efficiency with gains of $1.4$~$7.0\times$ and $1.5$~$146.1\times$, respectively, while maintaining only a modest accuracy drop. For CNNs, accuracy decreases by $0.1\%$~$3.1\%$ using the $L_2$ distance similarity, $0.1\%$~$3.4\%$ with the $L_1$ distance similarity, and $0.1\%$~$3.8\%$ when employing the Chebyshev distance similarity. For transformer-based models, the accuracy drop ranges from $1.4\%$ to $3.0\%$.

cs.AR↗

arXiv2024

Near-field Communications with Extremely Large-Scale Uniform Arc Arrays: Channel Modelling and Performance Analysis

In this letter, we propose a new conformal array architecture, called extremely large-scale uniform arc array (XL-UAA), to improve near-field communication performance. Specifically,under the non-uniform spherical wavefront channel model, we establish mathematical modeling and performance analysis for XL-UAAs. It is shown that XL-UAAs have larger direction-dependent Rayleigh distance and uniform power distance than the conventional XL uniform linear arrays (XL-ULAs). Moreover, a closed-form expression for the signal-to-noise ratio (SNR) is obtained, which depends on collective properties of XL-UAAs, such as the distance between the user and the array center,as well as the arc radius. In addition, we show that the asymptotic SNR of XL-UAAs with the number of antennas depends on the projection distance of the user to the middle of the arc array. Finally, numerical results verify that XL-UAAs achieve a higher SNR than XL-ULAs, especially at larger user incident angles.

cs.IT↗

arXiv2024

A Homogeneous Graph Neural Network for Precoding and Power Allocation in Scalable Wireless Networks

Deep learning is widely used in wireless communications but struggles with fixed neural network sizes, which limit their adaptability in environments where the number of users and antennas varies. To overcome this, this paper introduced a generalization strategy for precoding and power allocation in scalable wireless networks. Initially, we employ an innovative approach to abstract the wireless network into a homogeneous graph. This primarily focuses on bypassing the heterogeneous features between transmitter (TX) and user entities to construct a virtual homogeneous graph serving optimization objectives, thereby enabling all nodes in the virtual graph to share the same neural network. This ``TX entity'' is known as a base station (BS) in cellular networks and an access point (AP) in cell-free networks. Subsequently, we design a universal graph neural network, termed the information carrying graph neural network (ICGNN), to capture and integrate information from this graph, maintaining permutation invariance. Lastly, using ICGNN as the core algorithm, we tailor the neural network's input and output for specific problem requirements and validate its performance in two scenarios: 1) in cellular networks, we develop a matrix-inverse-free multi-user multi-input multi-output (MU-MIMO) precoding scheme using the conjugate gradient (CG) method, adaptable to varying user and antenna numbers; 2) in a cell-free network, facing dynamic variations in the number of users served by APs, the number of APs serving each user, and the number of antennas per AP, we propose a universal power allocation scheme. Simulations demonstrate that the proposed approach not only significantly reduces computational complexity but also achieves, and potentially exceeds, the spectral efficiency (SE) of conventional algorithms.

eess.SP↗

arXiv2024

Asynchronous Cell-Free Massive MIMO-OFDM: Mixed Coherent and Non-Coherent Transmissions

In this letter, we analyze the performance of mixed coherent and non-coherent transmissions approach, which can improve the performance of cell-free multiple-input multiple-output orthogonal frequency division multiplexing (CF mMIMO-OFDM) systems under asynchronous reception. To this end, we first obtain the achievable downlink sum-rate for the mixed coherent and non-coherent transmissions, and then provide a closed-form expression for the case with the maximum ratio precoding. Subsequently, an efficient clustering algorithm is proposed to group access points into different clusters with the same quantized phase shift in each cluster. Numerical results demonstrate that the mixed coherent and non-coherent transmissions can effectively improve the sum-rate of CF mMIMO-OFDM systems under asynchronous reception.

cs.IT↗

arXiv2024

SigDLA: A Deep Learning Accelerator Extension for Signal Processing

Deep learning and signal processing are closely correlated in many IoT scenarios such as anomaly detection to empower intelligence of things. Many IoT processors utilize digital signal processors (DSPs) for signal processing and build deep learning frameworks on this basis. While deep learning is usually much more computing-intensive than signal processing, the computing efficiency of deep learning on DSPs is limited due to the lack of native hardware support. In this case, we present a contrary strategy and propose to enable signal processing on top of a classical deep learning accelerator (DLA). With the observation that irregular data patterns such as butterfly operations in FFT are the major barrier that hinders the deployment of signal processing on DLAs, we propose a programmable data shuffling fabric and have it inserted between the input buffer and computing array of DLAs such that the irregular data is reorganized and the processing is converted to be regular. With the online data shuffling, the proposed architecture, SigDLA, can adapt to various signal processing tasks without affecting the deep learning processing. Moreover, we build a reconfigurable computing array to suit the various data width requirements of both signal processing and deep learning. According to our experiments, SigDLA achieves an average performance speedup of 4.4$\times$, 1.4$\times$, and 1.52$\times$, and average energy reduction of 4.82$\times$, 3.27$\times$, and 2.15$\times$ compared to an embedded ARM processor with customized DSP instructions, a DSP processor, and an independent DSP-DLA architecture respectively with 17% more chip area over the original DLAs.

cs.AR↗

arXiv2024

Power Allocation for Cell-Free Massive MIMO ISAC Systems with OTFS Signal

Applying integrated sensing and communication (ISAC) to a cell-free massive multiple-input multiple-output (CF mMIMO) architecture has attracted increasing attention. This approach equips CF mMIMO networks with sensing capabilities and resolves the problem of unreliable service at cell edges in conventional cellular networks. However, existing studies on CF-ISAC systems have focused on the application of traditional integrated signals. To address this limitation, this study explores the employment of the orthogonal time frequency space (OTFS) signal as a representative of innovative signals in the CF-ISAC system, and the system's overall performance is optimized and evaluated. A universal downlink spectral efficiency (SE) expression is derived regarding multi-antenna access points (APs) and optional sensing beams. To streamline the analysis and optimization of the CF-ISAC system with the OTFS signal, we introduce a lower bound on the achievable SE that is applicable to OTFS-signal-based systems. Based on this, a power allocation algorithm is proposed to maximize the minimum communication signal-to-interference-plus-noise ratio (SINR) of users while guaranteeing a specified sensing SINR value and meeting the per-AP power constraints. The results demonstrate the tightness of the proposed lower bound and the efficiency of the proposed algorithm. Finally, the superiority of using the OTFS signals is verified by a 13-fold expansion of the SE performance gap over the application of orthogonal frequency division multiplexing signals. These findings could guide the future deployment of the CF-ISAC systems, particularly in the field of millimeter waves with a large bandwidth.

cs.IT↗

arXiv2023

DeepBurning-MixQ: An Open Source Mixed-Precision Neural Network Accelerator Design Framework for FPGAs

Mixed-precision neural networks (MPNNs) that enable the use of just enough data width for a deep learning task promise significant advantages of both inference accuracy and computing overhead. FPGAs with fine-grained reconfiguration capability can adapt the processing with distinct data width and models, and hence, can theoretically unleash the potential of MPNNs. Nevertheless, commodity DPUs on FPGAs mostly emphasize generality and have limited support for MPNNs especially the ones with lower data width. In addition, primitive DSPs in FPGAs usually have much larger data width than that is required by MPNNs and haven't been sufficiently co-explored with MPNNs yet. To this end, we propose an open source MPNN accelerator design framework specifically tailored for FPGAs. In this framework, we have a systematic DSP-packing algorithm to pack multiple lower data width MACs in a single primitive DSP and enable efficient implementation of MPNNs. Meanwhile, we take DSP packing efficiency into consideration with MPNN quantization within a unified neural network architecture search (NAS) framework such that it can be aware of the DSP overhead during quantization and optimize the MPNN performance and accuracy concurrently. Finally, we have the optimized MPNN fine-tuned to a fully pipelined neural network accelerator template based on HLS and make best use of available resources for higher performance. Our experiments reveal the resulting accelerators produced by the proposed framework can achieve overwhelming advantages in terms of performance, resource utilization, and inference accuracy for MPNNs when compared with both handcrafted counterparts and prior hardware-aware neural network accelerators on FPGAs.

cs.AR↗

arXiv2021

Study on the Data Processing of the IOT Sensor Network Based on Hadoop Cloud Platform and TWLGA Scheduling Algorithm

The Internet of Things (IOT) sensor network is an effective solution for monitoring environment condition. IOT sensor network generates massive data, and the abilities of massive data storage, processing and query become technical challenges. To solve the problem, a Hadoop cloud platform is proposed. With the help of time and workload genetic algorithm (TWLGA), the data processing platform provides the work of one node to share with others, which not only raises efficiency of one single node, but also provides the compatibility support to reduce the possible risk of software and hardware. In the experiment, a Hadoop cluster platform with TWLGA scheduling algorithm is built, and the performance of the platform is tested. The results show that the Hadoop cloud platform is suitable for big data processing of the IOT sensor network.

cs.NI↗