Amazon said on its official blog on Oct. 19 that it will unveil TorchNeuron, which lets researchers and developers run PyTorch code on its Trainium AI chips without modifying it. Amazon said 85 percent of AI researchers and developers use PyTorch.

Until now, developers using PyTorch had to rewrite and validate code over months to move to a new chip, but TorchNeuron requires only a single settings change.

Developers can also keep using the testing, debugging and scaling tools they already use.

Amazon will open-source TorchNeuron at PyTorch Conference North America on Oct. 20.

Amazon said the biggest cost for many companies, including startups, is not chips but developers' time. It said product improvements are delayed by the time developers spend fixing software for new hardware.

Amazon also highlighted that if developers can switch chips easily, competition among chipmakers increases and development costs fall.

Amazon is also working with the PyTorch community to improve code so PyTorch is not tied to specific chips.

Cartesia, a voice AI company, is running inference on Trainium with almost no changes to its PyTorch code. The River Institute for Developmental Research, which studies treatments for neuropsychiatric disorders such as schizophrenia, is also using TorchNeuron for new drug development. Hugging Face provides default support for Trainium in its AI development tools.

TorchNeuron supports Trainium2 and Trainium3 chips.