Amazon releases TorchNeuron to run PyTorch on Trainium chips
Amazon announced a public beta of TorchNeuron, a software tool that lets PyTorch code run on its Trainium AI chips with only a basic configuration change. The tool has been in private beta since late last year and has been used by more than 30 companies and internal Amazon teams.
Key points
- TorchNeuron lets PyTorch code run on Trainium with only a configuration change, ending the need for custom interfaces.
- More than 30 companies and internal teams used TorchNeuron in private beta; Deepgram saw about 60% more audio per dollar on Trainium.
- Amazon is reportedly in talks to acquire Decart in a deal that could value the startup at about $7 billion.
The move is aimed at reducing Nvidia’s software advantage built around CUDA. Deepgram said its benchmarks showed Trainium processing about 60% more audio per dollar than its previous GPU setup. Inside Amazon, the Prime Video team used TorchNeuron to build Prime 360, an AI system that reconstructs NBA plays in 3D for select broadcasts. TorchNeuron will be released as open‑source software on Tuesday at the PyTorch Conference.
Amazon is reportedly in talks to acquire Decart, a startup that helps move models from Nvidia to alternatives, in a deal that could value Decart at about $7 billion. Amazon’s chip business, including Trainium and Graviton, reported an annual revenue run rate of more than $20 billion, which CEO Andy Jassy said could rise to about $50 billion as a standalone business.
The story so far
3 episodes →- Amazon releases TorchNeuron to run PyTorch on Trainium chipsthis story
Amazon challenges Nvidia’s software advantage with new tool for its AI chips
geekwire.com · 18 October 2026
Loading the full article…
This text was published by geekwire.com and written by Todd Bishop. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- Accenture expands federal AI push with Google's Gemini agents · 2 src
- DeepSeek launches Search and Fetch APIs at $0.75 per 1,000 requests · 4 src
- KDnuggets compares LangChain, LlamaIndex and building LLM apps from scratch · 1 src
- NorForge reports over 20 exposed Cairn panels likely linked to malicious intrusions · 2 src
- Microsoft outlines Project Perception to guard agentic insiders · 3 src
Comments
via GitHub Discussions