A Tensor Processing Unit (TPU) is a specialised AI chip developed by Google to accelerate the computations needed to train and run AI models. File
| Photo Credit: Reuters
Google has open-sourced a software stack that allows developers to run artificial intelligence workloads built using PyTorch on its Tensor Processing Units (TPUs) with minimal code changes, following a similar move by Amazon as cloud providers seek to reduce the software barriers to adopting alternatives to Nvidia’s AI chips.
A Tensor Processing Unit (TPU) is a specialised AI chip developed by Google to accelerate the computations needed to train and run AI models.
Called TorchTPU, the software enables developers to train, serve and deploy AI models on Google’s TPUs using familiar PyTorch application programming interfaces (APIs). Google announced the open-source release of TorchTPU at PyTorch Conference 2026 on October 20, alongside integrations with popular AI development and deployment tools.
The move targets a key advantage enjoyed by Nvidia, whose CUDA software platform, libraries and developer tools have helped establish a broad ecosystem around its graphics processing units (GPUs). By making their own chips easier to use with PyTorch, Google and Amazon are seeking to give developers more flexibility in choosing computing hardware without having to substantially modify existing code.
Amazon announced TorchNeuron, an open-source PyTorch backend for its Trainium2 and Trainium3 chips, two days back. Like Google’s TorchTPU, the tool is designed to let developers run supported PyTorch workloads on the company’s own accelerators with minimal changes.
Published – October 21, 2026 03:43 pm IST
