Machine learning model training. A large buffer of data waiting to be trained speeds up training. Especially with training on the GPU the biggest bottle neck is reading the data from disk.