Google and AWS Open New Paths for PyTorch Beyond NVIDIA
The New Stack, Thursday, October 22nd, 2026
Google's TorchTPU and AWS's TorchNeuron bring native PyTorch to TPUs and Trainium, though top performance still takes work.
At PyTorch Conference North America, Google open-sourced TorchTPU, a native PyTorch stack for TPUs built with Meta that removes JAX runtime dependencies, while AWS released TorchNeuron in public beta for Trainium2 and Trainium3.
Both aim to reduce dependence on NVIDIA's CUDA ecosystem.
Basic migration can need only small code changes, such as switching to the neuron device, and both support eager mode and torch.compile.
The New Stack cautions that neither is a silver bullet, and reaching top-tier performance on non-NVIDIA hardware still takes tuning.