According to the arXiv abstract, TorchTPU introduces a native and principled PyTorch backend for Google TPUs. The project is an open-source collaboration between Google and Meta, and its stated goal is to enable seamless out-of-the-box model execution and straightforward migration of PyTorch workloads to TPUs.
The architecture is described as hybrid, combining eager execution with compiled execution. This is intended to give PyTorch users a native path to TPU hardware while preserving familiar PyTorch semantics.
The abstract does not include specific benchmark results or performance comparisons, so the practical speedups and trade-offs of the hybrid design are not detailed in the source.