GPU-WORKLOAD NETWORK OPTIMIZATION FOR AI AND GEOSPATIAL ANALYTICS

Abstract

GPU-accelerated computing has become the dominant paradigm for artificial intelligence, machine learning, and geospatial analytics workloads. These workloads place unique and demanding requirements on the local area network. They generate large-volume data transfers between compute nodes, storage systems, and end-user workstations, with traffic patterns that differ fundamentally from those of conventional enterprise applications. This paper presents the network architecture designed and deployed for a national geospatial intelligence and AI analytics subsidiary of a regional technology group. The end-to-end infrastructure upgrade encompassed switches, wireless access, firewalls, and data centre network enhancements, with a specific focus on optimizing LAN performance for GPU-intensive workloads. Four key contributions emerge from this work. The first is a traffic characterization methodology for GPU computing environments. The second is a LAN architecture design that addresses the specific throughput, latency, and traffic pattern requirements of distributed GPU workloads. The third is a switch architecture optimization approach that maximizes performance for large file transfers and parallel data access. The fourth is a set of empirical performance observations drawn from the production deployment. Together these contributions provide practical guidance for network architects designing LAN infrastructure to support AI and analytics environments where GPU-accelerated computing is the primary workload.