Where Is VLA Inference Time Spent? Bottlenecks and Optimization Interactions Across GPUs

This paper presents a comprehensive measurement study of seven families of inference optimization methods—including CUDA graphs, compilation, token pruning, and quantization—evaluated on seven state-of-the-art Vision-Language-Action (VLA) models across four GPU families ranging from datacenter H100 to embedded Jetson Thor. The authors analyze how CPU launch overhead, memory bandwidth, and kernel efficiency jointly influence optimization gains and how applying one method affects the benefit of another. Based on this analysis, they derive VLA-Triage, a set of guidelines for stacking optimizations under latency budgets, identifying which stages can provide required savings and which combinations require testing in deployed execution modes.