Amazon Web Services announced TorchNeuron on November 4, a developer tool that, according to AWS, enables PyTorch applications to run on the company’s Trainium AI chip by changing a single line of configuration.
What TorchNeuron does
AWS says TorchNeuron is adapted from PyTorch, the open-source toolkit originally developed by Facebook and governed by the Linux Foundation since 2022, so developers familiar with PyTorch should be able to adopt the tool quickly. Amazon claims about 85% of AI applications are written with PyTorch.
The company states that TorchNeuron requires no code rewrite to move an app to Trainium. AWS argues that using Trainium on Amazon’s cloud gives startups access to the broader AWS infrastructure for moving from prototype to production, a point Colin Brace, VP of Technology Software at AWS Annapurna Labs, emphasized in emailed comments.
Performance, portability and platform strategy
AWS positions Trainium as a unified platform for both training and inference and says that combination improves economics across full workloads. Colin Brace pointed to Deepgram as an example, saying the company “moved to Trainium using native PyTorch support and they’re processing 70% more audio per dollar at production-grade accuracy.” Deepgram is an AI voice app provider and competitor to ElevenLabs.
The announcement frames TorchNeuron as part of a broader push to make PyTorch less tied to specific chips. AWS said it will contribute code upstream to the main PyTorch project to “remove code that ties the framework to specific chips” and to make PyTorch more hardware-agnostic, in the company’s words.
Amazon also contrasted its approach with so-called neocloud providers that are tied to Nvidia hardware, naming Coreweave, Nebius, and Firmus as examples. The company argues that building on AWS avoids depending on a single-purpose compute provider, a claim attributed to Colin Brace.
The tool’s release and AWS’s pledge to contribute changes to PyTorch signal an attempt to broaden hardware choices for developers while keeping PyTorch compatibility intact, according to the statements published by AWS on November 4.
Original source: TechCrunch AI