Google open-sources TorchTPU, letting PyTorch apps run on its TPU chips

Google announced Oct. 20 that TorchTPU, its native PyTorch stack for Tensor Processing Units, is now open source on GitHub.

Why it matters: Most AI apps are built with PyTorch and run by default on Nvidia GPUs. TechCrunch reported that Google and Amazon are both trying to loosen Nvidia's hold on AI developers.

What it does

Google said TorchTPU lets existing PyTorch workloads run on TPUs with minimal code changes. It was co-developed with Meta's PyTorch team. Google said it also works with Hugging Face, vLLM, SGLang, llm-d, TorchTitan, Ray and VeRL.

Google said TorchTitan training on TorchTPU reached up to 52% model FLOPs utilization on Qwen 3 8B. It said training scaled across up to 8,192 Ironwood TPUs.

Amazon's move

A day earlier, Amazon released TorchNeuron, which lets PyTorch apps move to its Trainium chips, TechCrunch reported. Google also launched a TPU agent suite in preview to help developers migrate.

Sources: Google Cloud blog; TechCrunch.