AI News · New apps ·
AWS launches TorchNeuron to move PyTorch apps onto Trainium chips
AWS released TorchNeuron, a tool that lets AI developers move PyTorch applications onto Amazon's Trainium chips with what Amazon says is a one-line configuration change. Amazon said about 85% of AI apps use PyTorch, where Nvidia's CUDA has made its hardware the default. AWS also said it will contribute code to the main PyTorch project to reduce dependence on specific chips.
Key points
- TorchNeuron moves PyTorch apps to Trainium with a one-line configuration change, Amazon says.
- Amazon says about 85% of AI apps use PyTorch.
- Deepgram reported processing 70% more audio per dollar on Trainium.
What happened: AWS released TorchNeuron on Monday, a developer tool that helps AI startups move PyTorch applications onto Amazon's Trainium chips, TechCrunch reported. Amazon says apps need only a one-line configuration change and no code rewrite.
Background: PyTorch is the open-source toolkit that Amazon says about 85% of AI apps use. Nvidia's CUDA software has made Nvidia hardware the default choice within PyTorch, which TorchNeuron is designed to challenge. Amazon said it will also contribute code to the main PyTorch project to reduce its dependence on specific chips, including Nvidia's. An AWS Annapurna Labs technology executive said: "The real story is removing the lock-in that kept developers from having a real choice at all."
The details: AWS pitches Trainium as handling both training and inference on one platform tied into its cloud infrastructure. It argues this is a better bet for startups than neoclouds such as CoreWeave, Nebius and Firmus, many of which are tied to Nvidia. AWS cited Deepgram, which reported processing 70% more audio per dollar on Trainium.
What to watch: TechCrunch noted that PyTorch exists to hide hardware complexity, so more hardware choice could also bring more configuration work for developers. Pricing details for TorchNeuron were not reported.
Our take
If the one-line switch holds up, teams running PyTorch workloads gain a practical way to compare Nvidia and Trainium costs without rewriting code.