Amazon announced TorchNeuron on October 19, a software tool intended to let developers run existing PyTorch artificial intelligence code on its AWS Trainium chips with a single configuration change. The company says the approach removes the need to rewrite code when moving to its hardware, lowering a barrier to adoption.
The announcement targets a key advantage of Nvidia, whose CUDA software helps AI developers use its processors. Reuters reported that Nvidia has spent more than a decade developing that software ecosystem. Google, which also designs its own AI chips, is seeking to attract developers with improved PyTorch tools as well.
According to Amazon, TorchNeuron connects through PyTorch's standard mechanism for registering new hardware. The company says this allows familiar testing, debugging and scaling tools to work on Trainium without modification. In an Amazon demonstration, the same training script ran on a conventional chip and on Trainium after changing the device setting.
Colin Brace, a software executive at Amazon's Annapurna Labs chip unit, told Reuters that more than 30 early customers were using the software across fields including robotics, world models and voice AI. He said the aim was to reduce the steps needed to try Amazon's processors and that switching could bring substantial savings. Those savings are a company claim, rather than a universal performance guarantee.
Amazon says it is also contributing changes to PyTorch to reduce dependencies on specific chips. Hugging Face has added native Trainium support to its development tools, according to Amazon. TorchNeuron also integrates with TorchTitan, providing a standard route for distributing model training across multiple Trainium instances.
The company plans to release TorchNeuron under the Apache 2.0 open-source licence at PyTorch Conference North America on October 20. It says the software supports Trainium2 and Trainium3. The release will give more developers access to the tool for evaluating whether Amazon's hardware suits their own AI workloads.
Comments