[DigitalToday reporter Chi-gyu Hwang (황치규)] Google has released an open-source project called TorchTPU that makes it easier to move PyTorch-based AI apps to its TPU chips and Google Cloud.

TechCrunch reported on Monday that TorchTPU is designed to enable existing PyTorch apps to run on TPUs without being rewritten. Developers can move apps to TPUs and Google Cloud by making simple configuration changes.

PyTorch, developed by Meta in 2016, is an open-source tool widely used to build, train and run AI models. Some estimates suggest 85 percent of all AI apps use PyTorch.

When Meta developed PyTorch, Nvidia GPUs were the standard hardware for AI research, so many PyTorch features were designed for Nvidia chips. Meta handed PyTorch over in 2022 to the Linux Foundation, a non-profit organisation.

Amazon earlier released its own PyTorch version, TorchNeuron, on Sunday. TorchNeuron supports moving PyTorch apps to Amazon's Trainium AI chips and AWS without rewriting code.

Amazon is also pushing work to separate various features from Nvidia chips by submitting code to the PyTorch project, TechCrunch reported.

Google said it worked directly with Meta, Hugging Face, the vLLM developer Inferact, and Sglang developer RadixArk, among others, to ensure TorchTPU works well with major AI development tools like existing PyTorch does.

Google also launched a TPU Agent Suite to help with app migration and bug fixes. It also updated LiteRT, which runs AI on smartphones, laptops and embedded devices, to better integrate with PyTorch apps.

It is hard to say that AI app developers will immediately abandon Nvidia hardware in large numbers based on these changes alone, but Google and Amazon are trying to reduce Nvidia's competitive edge by lowering barriers to switching to alternative chips and cloud services, TechCrunch reported.