Migrating existing PyTorch workloads from Nvidia GPUs to Google TPUs using TorchTPU typically requires changing `.to("cuda")` to `.to("tpu")`.