How do you choose between GPU, Inferentia and Trainium for training and inference?

GPUs offer the broadest framework support and are the default for flexibility; Trainium targets large-scale training and Inferentia targets inference, both offering materially better price/performance for supported transformer workloads at the cost of a Neuron compilation step and a narrower operator surface.

This is a price/performance versus flexibility trade, and the honest answer requires benchmarking rather than a rule.

GPU instances