Google has introduced TorchTPU, an open source project designed to help artificial intelligence developers run applications on Google’s Tensor Processing Units, or TPUs.

Announced on Oct 20, the technology represents another effort to reduce the AI industry’s dependence on Nvidia’s graphics processing units, which currently power a large proportion of AI development and deployment.

The initiative follows a similar move by Amazon, which recently introduced its own software to support applications running on its proprietary AI chips.

PyTorch is a widely used open source framework that enables developers and startups to build, train and run artificial intelligence models.

Originally developed by Meta in 2016, the framework became closely associated with Nvidia’s hardware because many of its features were designed to work with the company’s GPUs.

Meta transferred PyTorch to the Linux Foundation in 2022, helping establish a broader community to support its continued development and maintenance.

Google’s TorchTPU is designed to make it easier for developers to transfer existing PyTorch applications to Google’s TPU hardware and cloud infrastructure without rewriting their software.

Amazon has introduced a comparable solution called TorchNeuron, which allows developers to adapt applications for its Trainium AI chips and Amazon Web Services.

Amazon is also contributing code to the main PyTorch project to reduce the framework’s dependence on Nvidia specific features and improve compatibility with alternative hardware.

Google says it has worked with organisations including Meta, Hugging Face, Inferact and RadixArk to ensure that TorchTPU integrates effectively with popular AI development tools and libraries.

The company has also introduced a dedicated TPU agent suite to help developers migrate applications, identify errors and resolve technical problems during the transition.

In addition, Google has updated its LiteRT technology, which supports AI applications on smartphones, laptops and embedded devices, to improve compatibility with PyTorch based applications.

The introduction of TorchTPU does not necessarily mean that developers will immediately abandon Nvidia’s hardware, which benefits from an established ecosystem of software, tools and technical expertise.

However, Google and Amazon are making it easier for businesses to consider alternative chips and cloud platforms, potentially increasing competition across the AI infrastructure market.

If these tools prove reliable and cost effective, they could gradually weaken Nvidia’s advantage by giving developers greater flexibility in choosing the hardware used to build and operate their AI applications.

Source: TechCrunch