Google has launched TorchTPU, a specialized version of the PyTorch framework designed to enable AI applications to run on Google Tensor Processing Units. This move aims to reduce the industry's reliance on Nvidia GPUs, which currently support the vast majority of AI workloads.
The new framework allows developers to migrate existing AI models to Google hardware with minimal configuration changes, eliminating the need to rewrite underlying code. Google collaborated with Meta, Hugging Face, Inferact, and RadixArk to ensure compatibility with various developer tools.
In addition to TorchTPU, Google introduced a TPU agent suite to assist in migrating AI applications to Google Cloud by automatically fixing errors. The company also updated LiteRT to support PyTorch-based applications on smartphones, laptops, and embedded devices.
This development follows a similar announcement from Amazon Web Services, which recently introduced TorchNeuron to support its Trainum chips. Amazon also contributed code to the original PyTorch project to improve compatibility with non-Nvidia hardware.
