In this paper we evaluate three state-of-the-art neural-network-based approaches for large-scale video classification, where the computational efficiency of the inference step is of particular importance due to the ever increasing amount of data throughput for video streams. Our evaluation focuses on finding good efficiency vs. accuracy tradeoffs by evaluating different network configurations and parameterizations. In particular, we investigate the use of different temporal subsampling strategies, and show that they can be used to effectively trade computational workload against classification accuracy. Using a subset of the YouTube-8M dataset, we demonstrate that workload reductions in the order of 10×, 50× and 100× can be achieved with accuracy reductions of only 1:3%, 6:2% and 10:8%, respectively. Our results show that temporal subsampling is a simple and generic approach that behaves consistently over the considered classification pipelines and which does not require retraining of the underlying networks.