> Not compatible with pytorch either. PyTorch **is** compatible with TPUs, albeit you will need some minor modifications and it has a slightly clunky debugging toolset. You will have to use e.g. torch_xla.core.xla_model as xm and xm.optimizer_step(optimizer) instead of optimizer.step(), but otherwise it's pretty much entirely supported. (and some pytorch ops aren't 100% supported tbf) If you stick to standard layers (Conv, Linear, Transformer blocks), it’s fine. Tensorflow on the other hand works 100% and was even designed with TPUs in mind. **TLDR:** - Switching from nvidia to TPUs if all you know is PyTorch is probably 1-2 weeks of training - If you already know TensorFlow: it's zero learning, you literally just wrap your current code in a TPU strategy (few hours at most) - Ecosystem & infra switch is obviously harder, as it means switching providers, parts of your infra, etc.

---