Google Teams Up with Amazon to Challenge Nvidia’s Dominance in AI Application Development
On Tuesday, Google unveiled TorchTPU, its anticipated iteration of PyTorch, making it an open-source offering. This move positions Google alongside Amazon in the effort to challenge Nvidia’s dominance in the AI sector, where the majority of AI applications are generally designed to operate on Nvidia’s GPUs.
PyTorch is a widely used open-source framework that enables developers and startups to create, train, and deploy AI models. Reports suggest that as much as 85% of AI applications utilize PyTorch. Originally created by Meta (formerly Facebook) in 2016, PyTorch came into existence long before the emergence of large language models (LLMs). At its inception, Nvidia’s GPUs were already the preferred choice for AI research, influencing many of PyTorch’s features to function optimally with them. In 2022, Meta transitioned PyTorch to the oversight of the Linux Foundation, a nonprofit organization that manages several major open-source initiatives.
With TorchTPU, Google enables AI applications to operate on its proprietary AI hardware, the Tensor Processing Units (TPUs). The framework streamlines the transition of existing PyTorch applications to TPUs and Google Cloud, allowing developers to switch with minimal changes required to their code.
In a similar vein, Amazon announced its own adaptation of PyTorch, named TorchNeuron, on Monday. TorchNeuron facilitates the adaptation of AI applications to Amazon’s custom AI chip, Trainium, as well as to AWS without necessitating code rewrites. Furthermore, Amazon intends to amplify its efforts to diversify away from Nvidia’s control by contributing code to the main PyTorch repository that separates various functionalities from Nvidia hardware, thus promoting broader hardware compatibility.
However, Google may have an edge in certain aspects. The company claims to have collaborated closely with Meta, Hugging Face, Inferact (the developer behind vLLM), RadixArk (creator of SGLang), and other notable entities to ensure that the capabilities and libraries of TorchTPU integrate seamlessly with widely used AI development tools, similar to the original PyTorch.
Additionally, Google has introduced a specialized toolset, the TPU agent suite, designed to assist developers in migrating their projects to Google’s chips and cloud infrastructure while resolving any issues that arise. The company also enhanced its widely-used LiteRT tool, intended for running AI on mobile devices, laptops, and embedded systems, to provide improved functionality with PyTorch applications.
Despite these advancements, it is unlikely that AI application developers will swiftly abandon Nvidia hardware simply because alternatives are now more accessible. Nevertheless, the initiatives by Amazon and Google indicate a concerted effort to diminish Nvidia’s stronghold in the market.
