Google 与 Amazon 发布 TorchTPU 和 TorchNeuron 挑战 CUDA

Robert Hundt, distinguished engineer and vice president at AWS, introduces TorchNeuron at PyTorch Conference 2026 in North America. (Source: PyTorch)

PC본문_6월 Mouser (in Article)//

AD Manager | AD178089411676 //

// AD Manager | AD178089411676

Google and Amazon have introduced solutions aimed at reducing artificial intelligence (AI) applications' dependence on Nvidia.

At PyTorch Conference 2026, held in San Jose, California, on Oct. 20 local time, the two companies unveiled open-source frameworks called TorchTPU and TorchNeuron, respectively, designed to simplify the migration of existing AI applications from Nvidia graphics processing units (GPUs) to their own chips.

PyTorch is a framework for developing AI applications. It was developed in 2016 primarily by Meta's, formerly Facebook's, AI research laboratory, FAIR. Today, approximately 85% of AI application developers worldwide use it to build, train and run models, making it the de facto standard for AI application development.

However, FAIR originally designed PyTorch around Nvidia GPUs and the CUDA development platform, which had already become industry standards. As a result, PyTorch played a major role in establishing Nvidia's market dominance because migrating AI applications developed with PyTorch to other chips was difficult.

Google's TorchTPU and Amazon's TorchNeuron are expected to challenge this dominance of Nvidia's ecosystem.

TorchTPU performance from Google and Meta. (Source: MarkTechPost)

Google developed TorchTPU in collaboration with Meta. According to Google, more than 98% of existing PyTorch models can run on TPUs simply by changing the device string in the code to model.to("tpu"). The conversion takes anywhere from a few minutes to a few days. Google said it had enabled TPUs to operate as native devices on an equal footing with GPUs and CPUs at the core PyTorch level.

Amazon Web Services' (AWS) TorchNeuron similarly allows PyTorch-based AI models to run directly on AWS's Trainium AI chips. Like Google's TorchTPU, it requires developers to change only a few lines of device settings in their code, replacing Nvidia's cuda with AWS's neuron, rather than rewriting the entire application.

Data showing that most AI applications are developed based on PyTorch. (Source: Second Talent)

Amazon has released TorchNeuron as a public beta through GitHub under the Apache 2.0 open-source license. Beyond making the TorchNeuron code publicly available, Amazon is also contributing code directly to the main PyTorch open-source project.

Previously, migrating AI applications optimized for Nvidia's ecosystem to Google's or Amazon's proprietary AI chips took at least several days and, in more complex cases, several months. The new tools effectively remove these time and cost barriers.

As TorchNeuron, TorchTPU and similar tools gain adoption, dedicated AI chips are expected to compete primarily on their underlying performance, power efficiency and pricing. However, additional time will be needed to verify compatibility, performance and costs for individual models.