Getting GPUs fully utilized can be tricky when dealing with video data. In this talk I will explore this topic in depth and present some insights gained during a year of Deep Learning on videos at TwentyBN. Specifically, the problem is that such data may no longer fit into system-memory and we had to devise some efficient loading and decoding schemes to load data as quickly as possible. The talk will feature set of best practices and some open source recommendations that can help accelerate your deep learning process on video problems in practice. Importantly, I will discuss various options and ideas for data storage, including our own open source video and image format: GulpIO. I will discuss the effect that a good video codec has and how various codecs compare to each other. Also, I will explain how to properly benchmark programs that read from disc using a tool called nocache. Finally, I will present a set of benchmarks on real-world data under real-world conditions.