Google has released TorchTPU, an open-source version of the PyTorch AI toolkit that runs on its own Tensor Processing Unit chips, TechCrunch reports. The move follows Amazon's launch of TorchNeuron on Monday and is the latest attempt by cloud giants to loosen Nvidia's hold on AI development, since most AI apps are written to run on Nvidia graphics chips by default.
PyTorch, created by Meta and now managed by the Linux Foundation, is used by a large majority of AI apps. Google says developers can move existing PyTorch apps to its chips and cloud with a simple configuration change rather than a rewrite. It says it worked with Meta, Hugging Face and the makers of vLLM and SGLang on compatibility, and it has launched a TPU agent suite to help with migration. TechCrunch notes that developers are unlikely to abandon Nvidia overnight.
Source: TechCrunch