Latest ArticlesDisaggregated memory architectures are rapidly emerging as a significant trend for next-generation data centers, promising to enhance resource utilization and scalability. Existing system software often employs rapid context switching between lightweight threads to generate concurrent remote data access requests, thereby masking the microsecond-level latency inherent in disaggregated memory. However, this approach introduces substantial thread scheduling and communication overhead, consuming a significant amount of CPU resources that would otherwise be dedicated to computation and thus creating a new performance bottleneck. To address this challenge, a userspace asynchronous data path for disaggregated memory is proposed. The proposed data path focuses on data prefetching and eviction, eliminating the high-overhead multi-threaded concurrency model in favor of an asynchronous I/O mechanism to achieve the overlap of computation and network I/O in fine-grained tasks. By combining adaptive batching with dynamic thread management strategies, CPU overhead is minimized while ensuring the timeliness and concurrency of remote memory requests, thereby effectively hiding the latency of disaggregated memory access. A prototype system is implemented on top of AIFM and evaluated with four typical application workloads. Experimental results show that, compared with AIFM, the proposed approach reduces the CPU overhead of data prefetching and eviction by 73% while simultaneously increasing the average end-to-end application throughput by 38%.
Since 2012, the growth rate of intelligent computing has outpaced Moore’s Law, achieving an order-of-magnitude increase per year, which poses unprecedented challenges to the architecture of computer. Ultra-large-scale intelligent computing clusters have been proposed as national new infrastructure in “Report on the Work of the State Council” in 2026. Considering the characteristics of human intelligence, i.e “small computing power, large network”, the enabling technologies of optical transmission and switching with lower power consumption, and new architectures and mechanisms driven by technological innovation, we present ten preliminary judgments on overseas intelligent computing clusters development trends from four aspects: cluster architecture, computing nodes, network technology, and technical ecosystem. This paper aims to offer insights and contribute to the national implementation of ultra-large-scale intelligent computing clusters as new infrastructure.
Computility federations coordinate heterogeneous resources across independent providers while preserving administrative autonomy. This autonomy limits federation-level visibility, forcing gateways to route from coarse-grained and delayed performance summaries rather than real-time internal states. Under such weak and lagged observability, deadline-miss risk for latency-critical tasks can become structurally concentrated in a small number of domains and time windows, even when federation-wide averages remain moderate. To address this problem, we propose per-domain SLO feedback (PDSF), a low-dimensional closed-loop mechanism that updates bounded routing tolls from window-level summaries and biases snapshot-driven self-routing without modifying domain-local schedulers. Theoretical analysis shows that, under a weak monotonicity assumption on routing response, PDSF exhibits boundedness, forgetting, and local stability under delayed feedback. Trace-driven experiments on a multi-domain emulator using Google cluster workloads show that, compared with deployable baselines, PDSF significantly alleviates upper-tail overload severity, reduces worst-domain risk exposure, and protects latency-critical traffic without noticeable degradation to best-effort traffic. Sensitivity analysis further shows that snapshot staleness and routing concentration jointly determine the mechanism’s operating boundary and clarifies how stale measurements and aggressive routing response amplify risk while reducing the effectiveness of low-dimensional feedback.
AI computility integrates geographically distributed intelligent computing resources via high-speed networks, providing efficient and elastic infrastructure for large-scale intelligent applications. Multi-agent debate (MAD), as a distributed multi-expert collaborative paradigm, connects multiple large language model (LLM) agents through computility networks to engage in iterative discussions from distinct roles, effectively enhancing collaborative reasoning quality. However, when facing jailbreak attacks—a typical class of attacks targeting LLMs—the security of the collaborative reasoning paradigm in MAD remains unexplored. To address this gap, a jailbreak attack method targeting MAD is investigated under a semi-black-box threat model. A structured prompt-rewriting template is designed that systematically integrates four strategies—narrative encapsulation, role-driven escalation, iterative refinement, and rhetorical obfuscation—to exploit the interactive mechanisms of multi-agent debate for inducing harmful outputs. Experiments on four mainstream MAD frameworks with GPT-4o, GPT-4, GPT-3.5-turbo, and DeepSeek demonstrate that the proposed method increases the average harmfulness of generated content by 28.14% to 80.34%, with attack success rates reaching 80% in certain scenarios. Building on these findings, preliminary defense experiments are further conducted across the input, generation, and propagation stages, providing initial evidence for the feasibility of lightweight multi-stage protection in MAD systems. These findings expose potential security vulnerabilities within multi-agent debate, validating the necessity of developing robust and targeted defense strategies before the widespread deployment of multi-agent collaborative systems under AI computility.
The increasing parameter scale and structural complexity of deep neural networks (DNNs) pose significant challenges for their efficient deployment in cloud-edge-end collaborative computing architectures, where edge and terminal nodes must support real-time inference under stringent constraints on computation, storage, inference latency, and communication overhead. Although model quantization effectively reduces resource consumption, traditional approaches based on fixed rules or empirical heuristics exhibit notable limitations in accuracy preservation, training stability, and adaptability across diverse architectures and application scenarios, particularly under very low bit-width conditions (≤4 bit). To mitigate these issues, learning-based quantization incorporates learnable components and diverse supervisory signals, substantially improving accuracy preservation, optimization robustness, and hardware adaptability under low-bit conditions. In this paper, we systematically review representative studies on learning-based quantization, centered on two primary paradigms: post-training quantization (PTQ) and quantization-aware training (QAT). We summarize performance enhancement strategies including learnable parameter modeling, reconstruction and approximation optimization, auxiliary supervision, and learning-driven methods under data and hardware constraints. We further organize existing technical approaches and their interrelationships from the perspectives of quantization process stages and learning signal mechanisms, and analyze the applicability of different strategies under varying data conditions, model scales, and application constraints. Finally, we discuss future research directions in learning-based quantization from three aspects: error-analysis-based optimization of quantization parameters and learning signals, the trade-off between accuracy and resources in learning-based quantization, and a unified quantization framework across model architectures.
MoE (mixture of expert) architectures have become a widely recognized paradigm for scaling large language models (LLMs), and have been adopted by many frontier models deployed at the network edge. However, deploying MoE-based LLMs on resource-constrained edge devices remains challenging due to their massive parameter sizes, which often exceed the memory capacity of individual platforms. Existing in-situ MoE serving systems adopt expert swapping strategies that offload inactive experts to secondary storage and load them on demand. As LLMs support increasingly longer context windows with correspondingly longer user inputs, the prefilling phase faces a critical bottleneck: long input sequences activate nearly all experts, and the resulting full-scale expert swapping generates substantial I/O traffic, leading to prolonged TTFT (time to first token) latency. Typical edge environments, such as smart homes and smart factories, comprise multiple trusted interconnected devices, including tablets, laptops, and smart home appliances connected via local area networks. These devices collectively form a pervasive edge computility network (ECN), where each node carries underutilize computing power, memory capacity, and I/O bandwidth. This motivates us to harness the aggregated resources of the ECN through coordinated cross-node scheduling. This paper presents Sirius, a collaborative MoE inference system for edge computility networks that pools the collective resources across ECN nodes to accelerate the prefilling phase and reduce TTFT. Specifically, Sirius proposes three key techniques: 1) an interleaved sequence partitioning strategy that resolves workload imbalance in causal attention computation and maximizes communication-computation overlap; 2) an adaptive token scheduling mechanism that jointly considers computation overhead, swap latency, and expert buffer cache states to fully utilize collective I/O bandwidth while achieving load balance; 3) a speculative expert prefetching scheme that exploits the high similarity between adjacent-layer hidden states to overlap expert loading with computation, effectively hiding I/O latency. Experiments on real-world ECN platforming with Mixtral-8x7B and Qwen3-30B-A3B demonstrate that Sirius achieves 1.2–4.0 times inference throughput improvement over state-of-the-art baselines.
With the increasing popularity of deep neural network (DNN) applications, Chiplet-based accelerators have become critical infrastructure for supporting large-scale DNN inference services on server platforms. However, the real-time and dynamic characteristics of parallel multi-DNN workloads pose severe challenges to scheduling algorithms, while existing algorithms suffer from limitations in decision overhead, service level agreement (SLA) assurance, and dynamic adaptability. To address these issues, we propose MAOSA, a multi-agent reinforcement learning based online scheduling algorithm for multi-DNN workloads on multi-Chiplet accelerators, aiming to maximize the overall SLA satisfaction rate. The proposed algorithm partitions DNN tasks into DNN blocks and models them as agents. It utilizes the attention mechanism to generate scheduling strategies in parallel, which include execution order and PE-level resource allocation, thereby significantly reducing decision overhead. Furthermore, a cooperative reward function considering time slack is designed to optimize SLA, and a periodic unified preemption mechanism is introduced to adapt to dynamic load changes. Experimental results show that, compared with state-of-the-art algorithms, the proposed algorithm improves the SLA satisfaction rate by up to 18% and system throughput (STP) by 31.9%, reduces tail latency by 28.7%, and decreases scheduling execution time by a factor of approximately 11, effectively enhancing system performance and service quality.
Deploying large language models (LLMs) in mobile and edge computing environments faces multiple challenges, including limited on-device resources, scarce wireless bandwidth, and the continual evolution of cloud-side models. Although speculative decoding based edge-cloud collaborative inference can reduce end-to-end generation latency by using a lightweight draft model at the edge and a target model in the cloud for parallel verification, existing methods usually rely on a tightly coupled relationship between the two models. In practical systems, frequent updates of the cloud model often require repeated synchronization of the edge-side draft model, leading to substantial communication overhead, higher latency, and limited scalability. To address this issue, ECSpec is proposed as a communication-efficient edge-cloud collaborative inference framework for model-evolving scenarios. Its key idea is a shared backbone architecture that enables a single static edge-side draft model to remain compatible with a family of continuously evolving cloud-side target models, thereby avoiding repeated retraining or weight downloading at the edge and significantly reducing communication and maintenance costs. In addition, a channel-aware adaptive speculation mechanism is introduced to dynamically adjust the draft length according to real-time channel conditions and energy budgets, achieving a better tradeoff between inference efficiency and energy consumption. Experimental results demonstrate that ECSpec delivers more stable and efficient collaborative LLM inference under communication-constrained edge environments.
Video anomaly detection is of great practical value in the fields of intelligent video surveillance, public security, etc. With the advancement of deep learning, conventional pixel-based video anomaly detection methods gradually expose their vulnerability to noises, high computation cost, while skeleton-based approaches have emerged as a research hotspot due to their robustness and efficiency. This study provides a detailed explanation of the fundamental concepts, learning paradigms, and methodological workflows of human skeleton-based video anomaly detection methods, and systematically reviews and classifies the representative skeleton-based studies published in the recent five years into different categories, including prediction-based method, reconstruction-based method, hybrid reconstruction and prediction-based method, hybrid reconstruction and clustering-based method, as well as unclassified method. An in depth analysis is given to explore the underlying principles and innovations of each type of method. Moreover, this study summarizes existing benchmark datasets and evaluation metrics for skeleton-based video anomaly detection, as well as gives a comparative performance analysis of the mainstream methods across the benchmarks. Currently, skeleton-based anomaly detection methods face significant challenges in skeleton feature extraction, modeling and understanding of multi-person interaction, etc. Therefore, this study further proposes scientifically grounded future directions with regard to robust skeleton feature extraction algorithm, multimodal information integration and heterogeneous data fusion, etc., aiming at exploring the positive factors for constructing robust video anomaly detection framework, as well as analyzing the existing problems and applicable research directions of these series of video anomaly detection methods in different aspects, including adaptation ability, generalization ability, running efficiency, etc.
Multi-modal 3D object detection is a pivotal task in autonomous driving and environment perception. It achieves accurate localization and recognition of objects in 3D space by integrating data from multiple sensors, such as LiDAR (light detection and ranging) point clouds and camera images. To address the common challenges of large scale variations and insufficient multi-scale feature extraction in existing methods, we propose a multi-modal 3D object detection method based on dynamic scaling convolution. For this method, we design a dual-branch multi-modal fusion architecture, which consists of a VoxelNet-based 3D backbone and a novel dynamic scaling network. The two branches collaboratively process LiDAR point clouds and image-derived virtual point clouds to achieve cross-modal feature alignment and adaptive fusion. The core of the dynamic scaling network is a dynamic scaling module, which integrates 3D and 2D dynamic scaling convolution operators specifically tailored for multi-modal feature processing. This module adaptively adjusts the sampling locations and the number of valid sampling points within the receptive fields of 3D and 2D branches, thus effectively capturing structural and semantic variations of multi-scale objects. Extensive experiments on two benchmark datasets demonstrate the proposed method outperforms more than ten prevailing methods in terms of AP (average precision). Specifically, it achieves mAP3D (3D mean average precision) scores of 86.87%, 46.68%, and 68.39% for the car, pedestrian, and cyclist categories on KITTI test set, and achieves a mAP (mean average precision) of 71.8% on nuScenes test set. These results demonstrate the superior detection performance and strong generalization ability of the proposed method for multi-modal 3D object detection tasks.