edge deployment

This paper systematically reviews the evolution, technical challenges, and future directions of Edge Artificial Intelligence (Edge AI). Addressing core constraints—including resource scarcity, security and privacy risks, high energy consumption, and unreliable connectivity—the study conducts a rigorous literature analysis guided by the PRISMA framework. It proposes a multidimensional taxonomy encompassing deployment location, computational hierarchy, application scenarios, and hardware architectures. The review integrates emerging paradigms such as TinyML, federated learning, and neuromorphic computing, identifying three key advancement pathways: continual learning, cloud-edge collaboration, and trustworthy AI integration. Furthermore, it synthesizes enabling technologies—including model compression, domain-specific accelerators, efficient communication protocols, and secure computation mechanisms. The resulting comprehensive reference framework bridges theoretical rigor and practical engineering guidance, serving both researchers and practitioners in the Edge AI domain.

To address the urgent Quality-of-Service (QoS) requirements of IoT devices—particularly low latency, high bandwidth efficiency, and strong privacy and security—this paper conducts a systematic investigation of serverless edge computing. We perform a rigorous Systematic Literature Review (SLR) to construct the first comprehensive, multi-dimensional taxonomy covering architectures, QoS metrics, implementation mechanisms, application domains, and communication patterns—enabling structured cross-study comparison. Our analysis identifies six key open challenges, including function cold-start latency, secure isolation, and low-latency deployment. The resulting authoritative reference framework has already informed over 30 subsequent research projects in topic selection and positioning, thereby advancing standardization efforts and empirical research in serverless edge computing.

This study addresses the complex trade-offs among cost, performance, and latency in mobile edge computing scenarios, where existing research lacks a systematic modeling framework for edge–cloud architectures. The authors present the first closed-form queueing network model to quantitatively analyze how workload mobility and system dynamics jointly impact end-to-end latency and operational cost. Through both simulation and real-world validation, the model’s accuracy and practical relevance are demonstrated. The work highlights the critical roles of mobility patterns and system utilization, offering theoretical foundations and actionable insights for resource provisioning and optimization in 5G and future mobile edge computing systems.

This work addresses workflow scheduling for cyber-physical systems (CPS) in the edge–cloud continuum, where heterogeneous edge devices face coupled constraints on computation, communication, energy, and sensing/actuation capabilities. The objective is to minimize end-to-end latency under these multi-dimensional resource limitations. To this end, we formulate a Mixed-Integer Linear Programming (MILP) model that explicitly captures the heterogeneity of compute, network, energy, and CPS-specific resources. We further design an enhanced heuristic algorithm to efficiently solve the NP-hard scheduling problem. Compared to state-of-the-art approaches that neglect realistic multi-constraint interactions, our method achieves an average latency reduction of 13.54% on real-world use cases and up to 33.03% on synthetic workflows. The solution demonstrates both significant performance gains and strong scalability. This work provides a theoretically grounded, empirically validated framework and a practical methodology for intelligent workflow orchestration in resource-constrained edge–CPS environments.

Dynamic decision-making between local processing and edge offloading in edge computing remains challenging due to hardware heterogeneity (e.g., diverse accelerators), time-varying network conditions, and multi-tenant interference. Method: This paper proposes a unified performance model integrating M/G/1 and M/M/m queuing theories to jointly characterize these factors. It further introduces an adaptive offloading resource manager enabling low-latency, online offloading decisions. Contribution/Results: Evaluated on cross-platform AI workloads and real-world edge environments, the model achieves a mean absolute percentage error of only 2.2%. Compared to static offloading policies, it reduces end-to-end latency by up to 37%, significantly improving service efficacy and resource utilization in dynamic edge scenarios.

Low resource utilization and the absence of fine-grained performance–power co-analysis mechanisms hinder CPU-intensive stream processing in edge computing. To address this, we propose the first lightweight synthetic microbenchmark framework tailored for single-node edge clusters, systematically characterizing the nonlinear coupling among workload scale, CPU frequency, and power consumption/performance. Leveraging multidimensional parameter scanning, real-time power monitoring, and normalized performance analysis, we empirically identify— for the first time—the power–performance inflection points, enabling data-driven dynamic configuration recommendations. Experimental evaluation demonstrates that our framework achieves up to 37% power reduction at equivalent performance levels, or an average 2.1× throughput improvement under identical power budgets. This work fills a critical gap in fine-grained resource profiling and joint optimization benchmarking for edge environments.

To address the task offloading scheduling challenge in edge-cloud computing under resource constraints, this paper systematically surveys cooperative scheduling techniques that jointly optimize low latency, high quality of experience (QoE), and strong fault tolerance. We propose a fine-grained taxonomy based on algorithmic attributes—such as heuristic and metaheuristic approaches—and establish a cross-layer unified evaluation framework to enable precise alignment between scheduling strategies and application scenarios. By integrating QoS modeling, fault-tolerance mechanisms, and multidimensional comparative analysis, we synthesize a comprehensive, structurally coherent knowledge graph for edge-cloud task scheduling. Our key contributions include: (i) a novel classification paradigm grounded in algorithmic characteristics; (ii) a scalable, scenario-aware evaluation framework; and (iii) a systematic identification of open challenges and future research directions—thereby providing both theoretical foundations and practical guidance for designing adaptive scheduling algorithms.

This study addresses the multi-objective trade-offs among generation quality, energy consumption, latency, and memory when deploying large language models on edge devices. The authors construct a reproducible empirical evaluation framework to systematically analyze the energy efficiency, performance, and privacy characteristics of models ranging from 0.5B to 9B parameters on a real-world Android device (Samsung Galaxy S25 Ultra). Leveraging non-intrusive, fine-grained power monitoring and mixed-precision inference, they uncover a “quantization-energy paradox”: model architecture—not quantization strategy—dominates energy consumption. Notably, Mixture-of-Experts architectures disrupt conventional scaling–energy relationships, and medium-scale models such as Qwen2.5-3B emerge as the optimal choice, balancing high output quality with energy efficiency, thereby offering practical deployment guidelines for on-device intelligence.

This work addresses the challenge of jointly optimizing energy efficiency and end-to-end latency in low Earth orbit satellite edge computing, where the stochastic nature of communication and computation delays complicates reliable performance guarantees. To overcome the conservatism of traditional approaches such as the Chebyshev–Cantelli inequality, the authors propose a unified delay modeling framework that integrates parameter estimation with quantile regression to accurately characterize the full distribution of latency. Leveraging this data-driven model, they dynamically adjust GPU clock frequencies to minimize energy consumption while ensuring that end-to-end latency remains below 500 ms with at least 95% reliability. The approach demonstrates over 50% energy savings and is validated across diverse image processing tasks and hardware platforms, confirming its generality and effectiveness.

This work addresses the challenge of deploying large language models on edge devices, which is hindered by a lack of systematic understanding of inference latency and energy efficiency scaling across heterogeneous hardware (CPU/GPU/NPU). The authors propose QEIL, a unified framework that, for the first time, uncovers stable power-law scaling behaviors of Transformer models with respect to latency, energy consumption, and task coverage. Leveraging these insights, QEIL introduces three composite metrics and a safety-aware intelligent scheduler to enable coordinated optimization across heterogeneous accelerators from diverse vendors. Through formal modeling, computational orchestration, thermal management, fault-tolerant execution, and hardware health monitoring, QEIL significantly improves energy efficiency, reduces latency, and expands task coverage across five model families—while preserving model accuracy and ensuring system safety.

This work addresses the challenge of achieving effective load balancing for object detection across heterogeneous edge devices, where hardware disparities and dynamically varying workloads and scene complexities make it difficult to jointly optimize accuracy, latency, and energy consumption. To this end, the authors propose a multi-objective load balancing approach that employs a two-stage decision mechanism for coordinated optimization. In the first stage, candidate devices are filtered based on accuracy awareness; in the second stage, a weighted scoring function incorporating expected latency and energy consumption enables real-time scheduling. Experimental results demonstrate that, compared to baseline methods, the proposed scheme reduces end-to-end latency by 80% and energy consumption by 50%, while incurring no more than a 10% loss in accuracy, thereby significantly enhancing overall system efficiency.

This work addresses the challenge of simultaneously achieving low latency, high energy efficiency, high reliability, and adherence to multiple resource constraints in edge-hub-cloud collaborative IoT workflow scheduling. To this end, it proposes the first exact multi-objective, multi-constraint joint optimization model based on Continuous-Time Mixed-Integer Linear Programming (CT-MILP). By explicitly modeling task dependencies and incorporating a selective task replication mechanism, the model enhances system reliability while minimizing redundant overhead. Evaluated on both real-world and synthetic workflows, the proposed approach outperforms state-of-the-art heuristic algorithms, reducing average latency by 29.83% and energy consumption by 33.96%, while improving reliability by 28.49%. The method demonstrates practical runtime efficiency and favorable scalability, making it suitable for complex IoT environments.