This is also not quite right. Models whose parameters are too big to fit on one GPU can be trained by splitting them across multiple GPUs, as was done here, for example: http://papers.nips.cc/paper/5346-sequence-to-sequence-learni...