Skip to content
Tech · Seattle

Amazon opens TorchNeuron to all developers, taking aim at Nvidia's software advantage

The tool runs standard PyTorch code on Amazon's Trainium chips after a basic configuration change.

·
The Amazon Spheres in downtown Seattle seen from above
The Amazon Spheres at the company’s Seattle headquarters.

Amazon said Monday it is opening TorchNeuron, software that runs code written in the PyTorch framework on its Trainium AI chips, to all developers as a public beta, and plans to release it as open source on Tuesday at the PyTorch Conference in San Jose. Until now, developers had to modify their code to call Trainium-specific interfaces.

“If you know PyTorch, then you know Trainium,” said Colin Brace, vice president of technology software at Annapurna Labs, Amazon’s chip-design unit. Nvidia’s lead in AI chips rests partly on software built around its CUDA platform, and adapting that code to other hardware can take months.

More than 30 companies and internal teams used TorchNeuron during a private beta. San Francisco speech AI company Deepgram says Trainium processed about 60% more audio per dollar than its GPU setup, and Prime Video used it to build Prime 360, which reconstructs NBA plays in 3D. Amazon tests the tool nightly against more than 100,000 models from Hugging Face.

CEO Andy Jassy wrote in April that Amazon’s chip business, including Trainium and Graviton, had an annual revenue run rate above US$20 billion.

From ARQS Sponsored

Readers looking to act on this kind of technology shift can start with Design Workbench, TradeHound, Website audit and Flow automation. In the ARQS Store, Wine Cellar Ledger — 1 Month Access, ARQS ContractDiff Desk, SponsorKit — Creator Media Kit & Rate Card Builder are among the newest tools available.