Sundar Pichai and Tim Cook
Source: Reuters, Apple
Apple said on Monday that the artificial intelligence models underlying its AI system, known as Apple Intelligence, are pre-trained on processors designed by Google, a sign that big tech companies are looking for alternatives to Nvidia when it comes to training cutting-edge AI.
Apple’s choice of Google’s home-grown Tensor Processing Unit (TPU) for training was detailed in a technical paper the company published earlier this week. Separately, Apple released a preview of Apple Intelligence on select devices on Monday.
NVIDIA’s expensive graphics processing units (GPUs) dominate the market for high-end AI training chips and have been in such high demand over the past few years that it has been difficult to procure the quantities needed. OpenAI, Microsoft, and Anthropic all use NVIDIA GPUs in their models, and other tech companies such as Google, Meta, Oracle, and Tesla are also buying GPUs to build their AI systems and services.
Meta CEO Mark Zuckerberg and Alphabet CEO Sundar Pichai both made comments last week suggesting their companies and others in the industry may be overinvesting in AI infrastructure, but also acknowledged that the business risks of not doing so are too high.
“The downside of falling behind is that you lose ground on the most important technologies of the next 10 to 15 years,” Zuckerberg said on a podcast with Bloomberg’s Emily Chang.
Apple doesn’t name Google or Nvidia in the 47-page paper, but it does say that the Apple Foundation Model (AFM) and AFM server are trained on “Cloud TPU clusters,” meaning Apple rented servers from cloud providers to run the calculations.
“This system enables efficient and scalable training of AFM models, including on-device AFM, AFM servers, and larger scale models,” Apple said in the paper.
Representatives for Apple and Google did not respond to requests for comment.
Apple revealed its AI plans later than many other companies, who have been vocal proponents of generative AI since OpenAI released ChatGPT in late 2022. On Monday, Apple unveiled Apple Intelligence, a system that includes several new features, including a revamped look for Siri, improved natural language processing, and AI-generated summaries in text fields.
Over the next year, Apple plans to roll out features based on generative AI, including image generation, emoji generation, and an improved Siri that can use a user’s personal information to take actions within apps.
Apple said in a paper on Monday that the on-device AFM was trained on a single “slice” of 2,048 TPU v5p chips working together — its most advanced TPU, first released in December — and that the AFM server was trained on 8,192 TPU v4 chips configured to work together as eight slices across a datacenter network, according to the paper.
According to Google’s website, the company’s latest TPUs can be reserved for three years for less than $2 per hour of chip usage. Google first introduced TPUs for internal workloads in 2015 and made them generally available in 2017. Today, TPUs are one of the most mature custom chips designed for artificial intelligence.
Still, Google remains one of Nvidia’s major customers: It uses Nvidia’s GPUs as well as its own TPUs to train its AI systems, and also sells access to Nvidia’s technology on its cloud.
Apple has previously said that inference — which means taking pre-trained AI models and running them to generate content or make predictions — would be performed in part on Apple’s own chips in its data centers.
This is Apple’s second technical paper on its AI systems, following a more general version published in June, when Apple said it was using the TPU to develop its AI models.
Apple is scheduled to report its quarterly earnings after the close of trading on Thursday.
Video: How generative AI’s massive power consumption is straining the power grid
