Sundar Pichai and Tim Cook
Source: Reuters; Apple
Apple said Monday that the artificial intelligence models that power Apple Intelligence, the company’s AI system, were pre-trained on processors designed by Google, a sign that major tech companies are looking for alternatives to Nvidia when it comes to training advanced AI.
Apple’s choice of Google’s own Tensor Processing Unit (TPU) for training was detailed in a technical paper the company just published. Separately, Apple released a preview version of Apple Intelligence for some devices on Monday.
Nvidia’s expensive graphics processing units (GPUs) dominate the market for high-end AI training chips and have become so sought after in recent years that they have become difficult to obtain in the quantities required. OpenAI, Microsoft and Anthropic all use Nvidia’s GPUs for their models, while other tech companies including Google, Meta, Oracle and Tesla are buying them up to bolster their AI systems and offerings.
Meta CEO Mark Zuckerberg and Alphabet CEO Sundar Pichai both said last week that their companies and others in the sector may be overinvesting in AI infrastructure. But they acknowledged that the business risk of doing otherwise is too high.
“The downside of being behind is that you’re not positioned for the most important technology of the next 10 to 15 years,” Zuckerberg said in a podcast with Bloomberg’s Emily Chang.
Apple doesn’t name Google or Nvidia in the 47-page report, but it does note that the Apple Foundation Model (AFM) and the AFM server were trained on “Cloud TPU clusters.” That means Apple rents servers from a cloud provider to perform the calculations.
“This system allows us to train AFM models efficiently and scalably, including AFM on-device, AFM server, and larger models,” Apple said in the document.
Representatives for Apple and Google did not respond to requests for comment.
Apple is reportedly unveiling its AI plans later than many of its rivals, which loudly embraced generative AI soon after OpenAI launched ChatGPT in late 2022. On Monday, Apple introduced Apple Intelligence, a system that will introduce several new features, including a new look for Siri, better natural language processing and AI-generated summaries in text fields.
Apple plans to roll out features based on generative AI in the coming year, including image generation, emojis and a more powerful Siri that can access a user’s personal data and perform actions within apps.
In Monday’s paper, Apple said AFM was trained on-device on a single “slice” of 2,048 TPU v5p chips working together. That’s the most advanced TPU, which was first launched in December. AFM server was trained on 8,192 TPU v4 chips configured to work together as eight slices over a data center network, the paper said.
Google’s latest TPUs cost less than $2 per hour of use if booked three years in advance, according to Google’s website. Google first introduced its TPUs in 2015 for internal workloads and made them publicly available in 2017. They are now among the most mature custom chips designed for artificial intelligence.
Still, Google is one of Nvidia’s top customers. It uses Nvidia’s GPUs, its own TPUs for training AI systems, and also sells access to Nvidia’s technology on its cloud.
Apple previously said that inference, which uses a pre-trained AI model to generate content or make predictions, would happen in part on Apple’s own chips in its data centers.
This is the second technical document about Apple’s AI system, following a more general version published in June. At the time, Apple said it was using TPUs in the development of its AI models.
Apple is expected to report its quarterly results after the market closes on Thursday.