Amazon announced the launch of the open-source tool TorchNeuron on the local time of the 19th. Its key feature is straightforward: developers can use the PyTorch code they have written directly; no need to rewrite it. By changing just one configuration, the code can run natively on Amazon’s own Trainium chip. This tool will be officially released under an open-source license at the North American PyTorch Conference in San Jose, California, on the 20th.
To prove that “seamless” isn’t just talk, Amazon engineers ran the same AI training script on traditional chips and on Trainium. The only difference was the setting parameters specified in the code for each chip. As a result, in both environments, the model learned exactly the same content and had the same learning speed. This means that the tools that developers use for testing, debugging, and expanding models can be directly transferred to Trainium without any changes.
Targeting CUDA monopoly, giving developers "freedom to switch cores"
The opponent of this strategy is NVIDIA’s 10-plus-year-old CUDA moat. PyTorch is led by Meta and adopted by the majority of AI developers, and NVIDIA has made it efficient to run PyTorch code on its own chips using CUDA, firmly integrating the ecosystem into GPUs. What Amazon wants to leverage is precisely this bond.
Amazon calls TorchNeuron a “long-overdue key element” in the industry – a truly viable choice of options. The logic is: when developers can freely switch between different chips without additional costs, chip manufacturers can only offer more competitive products to win market share; in turn, developers benefit from lower costs and faster iteration. Furthermore, Amazon is actively participating in the upstream development of PyTorch, collaborating with the community to remove the code that binds the framework to specific chips. The goal is to make PyTorch more “hardware independent” for the entire developer community, rather than serving only Trainium users.
In the midst of intense competition, Amazon’s move is both a step toward developing its own Trainium and a way to break up Nvidia’s monopoly on the open-source ecosystem. As “core replacement” transforms from a major issue into a parameter for change, the one-sided situation of GPU dominance begins to crack open a gap that can be exploited.