Vision language model deployment trade offs for drones with network and image size
Cloud, Edge, or Split? Profiling Onboard and Split Vision-Language Model Deployment for Drone AI
Distributed, Parallel, and Cluster Computing
Summary
Using smart computer programs to help drones understand pictures and instructions is tricky because drones have limited power and computing ability. Sending all the work to the cloud can cause delays and depends on good internet, while doing everything on the drone uses too much battery. The authors studied a split method that shares the work between drones and remote servers, and compared it to the other methods. They found that the best choice depends on the quality of the network and the size of the images the drone processes.
What this means in practice
- •For drone developers: Choose how to deploy vision-language models on drones by balancing onboard processing, cloud, or split setups based on network and image resolution constraints.
- •For network architects: Adjust network resources and capacity for drone AI tasks by understanding the trade-offs of cloud versus split model inference under varying conditions.
Authors
Zoha Azimi, Reza Farahani, Schahram Dustdar, Christian Timmerer
Abstract
Vision-Language Models (VLMs) enable edge devices like unmanned aerial vehicles (UAVs) to interpret visual observations and reason about complex environments using natural-language instructions. However, their practical deployment remains challenging as onboard inference is constrained by limited computational, memory, and energy resources, whereas cloud-based inference introduces communication latency, bandwidth overhead, and dependence on network connectivity. To address these limitations, split computing offers a promising alternative by partitioning VLM inference between the resource-constrained UAVs and more capable remote servers. However, the performance trade-offs among fully onboard, cloud-based, and split-computing architectures for lightweight VLMs have not yet been systematically profiled. This paper benchmarks these three deployment paradigms using SmolVLM-256M as a representative lightweight VLM. We quantify their inference latency, computational resource utilization, communication overhead, and energy consumption across varying image resolutions and network conditions. Our results show that no deployment strategy is universally optimal; instead, the preferred strategy depends on the interaction between network conditions and input image resolution.