Google opens TorchTPU to pull PyTorch developers off CUDA

Google released TorchTPU, an open-source PyTorch build for its Tensor Processing Units, one day after Amazon shipped its own PyTorch port for the Trainium chip. The target is a framework "used by 85% of the AI apps," according to TechCrunch, whose features were originally built around Nvidia's GPUs.

"Developers can do this with a simple configuration change and no re-writing of their apps," the report says. Google worked with Meta, Hugging Face, Inferact, and RadixArk on compatibility, and bundled a TPU agent suite to migrate apps and resolve errors. An updated LiteRT platform extends the PyTorch path down to phones, laptops, and embedded devices.

Amazon's piece, announced Monday, is the more intrusive one. Beyond TorchNeuron for Trainium, Amazon is "submitting code to the main PyTorch project that decouples various features from Nvidia's chips." PyTorch was created by Meta in 2016 and handed to the Linux Foundation in 2022, so Nvidia's influence sits in a project neither cloud owns.

TechCrunch does not oversell it. "None of this means AI app developers will automatically ditch Nvidia's hardware en masse just because they can now easily move," the piece reads. We've logged 201 Nvidia stories in the past 90 days; this is the second PyTorch port to non-Nvidia silicon in two days, after Amazon's TorchNeuron drop on Monday.