Only when you have a gradient-descent problem. Pytorch and Tensorflow are the only candidate frameworks to use here. When using Horovod, start with single node, multi-GPU and measure training performance. If this is not sufficient, look at a multi-node, multi-GPU setup.
