Google has made TorchTPU, its long-awaited version of PyTorch, available as an open source project, joining Amazon in efforts to loosen Nvidia’s dominance over AI development, where most applications are built by default to run on Nvidia GPUs.
PyTorch, the widely used open source toolkit for building, training and running AI models, was developed by Meta in 2016 and later handed to the Linux Foundation. Because Nvidia GPUs were already the standard for AI research, many of its features were designed around them. TorchTPU enables AI applications to run on Google’s Tensor Processing Units and move to Google Cloud through a simple configuration change, without rewriting code.
The release follows Amazon’s announcement of TorchNeuron, which lets AI apps run on its Trainium chips and AWS. Amazon is also contributing code to the main PyTorch project to separate features from Nvidia hardware.
Google said it worked with Meta, Hugging Face, Inferact and RadixArk to ensure compatibility with popular AI tools. It also introduced a TPU agent suite to help developers migrate and troubleshoot applications, and updated its LiteRT tool to better support PyTorch apps on phones, laptops and embedded devices.
