Google joined hands with Meta to keep Nvidia in check in the artificial intelligence (AI) semiconductor market.
According to Reuters on the 17th, Google said it has begun developing technology to optimize PyTorch, an open-source software (SW) for running AI chips developed by Meta, for its own AI chip, the tensor processing unit (TPU).
PyTorch is a tool that handles programming for running AI chips on behalf of developers. It is used by AI developers worldwide and is virtually a standard tool, but it is currently optimized for Nvidia chips. Because of this, developers had to learn how to use new tools to run chips other than Nvidia's, which became an obstacle for Google as it sought to expand the TPU ecosystem.
To solve this, Google launched an internal project, TorchTPU, to ensure PyTorch works smoothly with TPUs. Once the project is completed, developers will be able to keep using the PyTorch they already use and switch only the hardware from Nvidia chips to Google chips. To draw more developers into its chip ecosystem, Google is also considering opening parts of the technology that boosts compatibility with PyTorch.
In particular, Meta, which develops and manages PyTorch, is said to be working closely with Google's TorchTPU project, according to sources. Meta was recently reported to be in talks with Google to adopt TPUs worth several billion dollars. If the two companies cooperate, Google can increase its cloud service market share and AI chip sales, while Meta, which has declared the development of the next-generation AI "superintelligence," can save infrastructure expense. The two companies, fierce rivals in the online advertising market, have formed an alliance to keep Nvidia in check.
A Google Cloud Spokesperson said, "We are focused on providing the flexibility and scalability needed regardless of the hardware chosen by developers."