A GPU instance is recommended for most deep learning purposes. Training new models will be faster on a GPU instance than a CPU instance. You can scale sub-linearly when you have multi-GPU instances or if you use distributed training across many instances with GPUs. To set up distributed training, see

