Google open-sources TorchTPU

Google’s TorchTPU is a PyTorch backend for running models and operations on Google’s Tensor Processing Units. The open-source project includes TPU-specific kernels, a compilation cache, support for distributed training, and examples covering models such as Llama, Qwen and ResNet. It also supports torch.compile() for graph-mode execution. The project drew 124 points and 37 comments on Hacker News, and you can review the TorchTPU project.

For a business choosing AI infrastructure, this is a new option to track, not proof that switching to TPUs will cut costs or improve performance. The project’s documentation says its interactive notebooks can be run from a TPU VM, and its installation and build instructions involve specific tools and configuration. Teams assessing hardware and software choices can watch how the project develops and decide whether it fits their existing workloads and skills.