Apple Used Google’s Chips Instead of Nvidia to Train AI Models

Apple utilized Google’s tensor processing units (TPUs) rather than Nvidia’s GPUs to develop key components of its artificial intelligence (AI) infrastructure, according to a recent research paper published by Apple (via Reuters).

Nvidia holds approximately 80% of the AI chip market, which includes processors made by Google, Amazon, and other cloud computing firms. Despite Nvidia’s stronghold, Apple opted for Google’s TPUs, marking a significant decision in the tech landscape.

The research paper did not explicitly exclude the use of Nvidia chips but notably omitted any mention of them, highlighting Google’s TPUs instead.

Apple’s paper detailed the use of two types of Google’s TPUs organized into extensive chip clusters. To train its AI models for iPhones and other devices, Apple utilized 2,048 TPUv5p chips. For server-based AI models, Apple employed 8,192 TPUv4 processors.

Google offers access to its TPUs through the Google Cloud Platform, requiring customers to develop software within its ecosystem. This contrasts with Nvidia’s approach, which sells its GPUs and systems as standalone products.

The paper also hinted at the potential for developing even more sophisticated AI models using Google’s TPUs, suggesting a long-term collaboration between the two tech giants.

Apple’s research paper coincided with the rollout of its AI features to beta users this week. These new features, unveiled at Apple’s June developer conference, include the integration of OpenAI’s ChatGPT technology into Apple’s software.

The decision to disclose the extent of its reliance on Google’s TPUs in the recent paper comes after initial reports by Reuters in June, which noted Apple’s use of Google’s chips but did not detail the full scale.