Amazon Web Services on Monday released TorchNeuron, a tool that lets artificial intelligence applications written in PyTorch run on Amazon’s Trainium chips by changing one line of configuration, with no code rewrite, the company said. PyTorch, originally developed by Facebook and maintained by the Linux Foundation since 2022, is the toolkit Amazon says about 85 percent of AI applications are written with.
The strategic point is Nvidia’s software advantage. Nvidia spent years making its CUDA software layer the default path for PyTorch, so its hardware became the default choice. TorchNeuron is adapted from PyTorch itself, so developers who know one know the other, and Amazon says it will contribute changes upstream to the main PyTorch project to remove code that ties the framework to specific chips, making it more hardware agnostic for all developers, not only Trainium users.
Amazon’s commercial argument is integration. Trainium handles both training and inference on one platform inside AWS, so a startup gets the surrounding cloud infrastructure without assembling it around a single purpose compute provider. “The real story is removing the lock-in that kept developers from having a real choice at all,” Colin Brace, vice president at AWS Annapurna Labs, told TechCrunch. AWS cited voice AI firm Deepgram, which it said processes 70 percent more audio per dollar after moving to Trainium, a company claim.
Whether uncoupling PyTorch from Nvidia gives developers cheaper choice or new configuration headaches is untested. AMD and Apple chips can already be targeted from PyTorch with a configuration change, so Amazon is not first; its claim is scale and cloud integration.