Amazon unveiled TorchNeuron, a native PyTorch backend for the AWS Trainium2 and Trainium3 AI chips. The company says existing PyTorch code runs on Trainium by switching the device to 'neuron', with eager execution, torch.compile and distributed training supported. The code is due to be released as open source at PyTorch Conference on 20 October. Firms including Hugging Face, Deepgram and Cartesia are already using it.

What it means: Nvidia's biggest advantage is not just its chips but the software around them. Most teams have written their models in PyTorch on the assumption they will run on Nvidia cards, and moving to other hardware has so far meant a fair amount of rewriting. With TorchNeuron, Amazon is trying to remove that switching cost by making Trainium look like just another device inside a familiar environment. If the promise holds in practice, businesses gain real choice on price and availability, which matters at a time when demand for compute outstrips supply. What still has to be shown is performance on production workloads, not only compatibility. For Greek teams training or serving models in the cloud, the practical lesson is to write device-agnostic code so they can compare costs across different hardware. Releasing the code as open source will also let the community judge how mature the stack really is. Competition at the software layer could end up lowering prices for everyone.

Read the original on About Amazon

Related service: Custom software

Related stories