Amazon Hits Nvidia's Monopoly With New TorchNeuron Tool

Amazon Web Services (AWS), Amazon's cloud division, has announced TorchNeuron. It is a library that lets PyTorch applications run on Amazon's own AI chips, Trainium, with no changes needed. PyTorch is an open source machine learning framework. Developers use it to build and deploy AI models. AWS says 85% of all AI applications are built on PyTorch. Switching to Trainium takes just one line of configuration. Until now, Nvidia had effectively locked up PyTorch support through its CUDA software stack. CUDA is what lets AI applications run on Nvidia GPUs. After moving to Trainium, the speech recognition platform Deepgram was able to process 70% more audio per dollar spent. It kept production-level accuracy the whole time.

This is really a fight between chipmakers and cloud providers, not a debate about which AI algorithm is better. Developers usually never see which chip is doing the work under the hood. AWS says it is also open-sourcing part of its changes and contributing them back to the main PyTorch repository. The goal is to decouple PyTorch from any single hardware vendor. That could help chip away at Nvidia's dominance in AI software and give developers more options for hardware acceleration.