6G Network Slicing for Low-Latency AI-Edge Deployments

Abstract

The evolution toward sixth-generation (6G) wireless networks represents a paradigm shift in how connectivity, computation, and intelligence converge to enable a new class of applications. At the forefront of this evolution are low-latency AI-edge deployments—scenarios in which artificial intelligence (AI) inference and decision-making occur at or near edge devices, necessitating end-to-end communication delays in the sub-millisecond range. Traditional 5G network slicing approaches, which statically or reactively allocate resources, often struggle to meet these stringent latency requirements under highly dynamic traffic patterns. This manuscript presents a comprehensive investigation of advanced 6G network slicing strategies tailored specifically for low-latency AI-edge use cases. We begin by contextualizing 6G’s technical enablers—such as terahertz (THz) communications, reconfigurable intelligent surfaces (RIS), and integrated sensing, computing, and communication (ISCC)—and their implications for supporting ultra-reliable, low-latency communications (URLLC). Next, we review the state of the art in network slicing, highlighting the limitations of static and reactive paradigms and the promise of AI-driven predictive orchestration. We then describe our dual-pronged methodological framework, which blends analytical latency modeling with large-scale, event-driven simulation to evaluate slice instantiation times, queuing delays, and end-to-end packet latency across three distinct slicing schemes: static, reactive, and predictive. A detailed statistical analysis—comprising thousands of inference request samples under Poisson and bursty arrival processes—reveals that predictive slicing reduces mean end-to-end latency by approximately 35% relative to static slicing and by 11% relative to reactive slicing, while also tightening latency variance and 95th percentile tail behavior. Importantly, we quantify orchestration overheads and resource utilization efficiencies, demonstrating that the modest control-plane cost of predictive models is offset by substantial improvements in latency and resource elasticity.