Apple Leverages Google’s Custom Chips for AI Training

In a recent development, Apple disclosed that its artificial intelligence models underpinning Apple Intelligence were pretrained using Google-designed processors. This strategic choice highlights a significant shift as major tech companies seek alternatives to Nvidia’s widely-used GPUs for AI training.

According to a technical paper released by Apple, the AI models were trained on Google’s Tensor Processing Units (TPUs). These specialized chips, originally created for Google’s internal workloads, are now being adopted more broadly. Apple’s decision to utilize Google’s TPUs suggests a trend among tech giants to diversify their hardware for AI training, moving beyond the traditional reliance on Nvidia.

Apple’s document does not explicitly name Google or Nvidia but specifies that its AI Foundation Model (AFM) and AFM server were trained using “Cloud TPU clusters.” This indicates that Apple utilized rented servers from a cloud service provider for the extensive computations required for AI training.

The company outlined in the paper that it employed 2,048 of Google’s TPUv5p chips for training its on-device AI model and 8,192 TPUv4 processors for its server-based AI model. Google’s TPUs, available through its cloud platform, offer a cost-effective solution for high-end AI training tasks, priced under $2 per hour when booked for three years in advance.

Apple’s move to integrate Google’s TPUs reflects a broader trend in the industry. Nvidia’s GPUs have been in high demand due to their effectiveness in AI training, but the supply has struggled to meet the growing needs of companies like OpenAI, Microsoft, and Anthropic. Meanwhile, firms like Google, Meta, Oracle, and Tesla have been expanding their AI capabilities, often resorting to custom solutions like TPUs.

Apple’s unveiling of Apple Intelligence, which includes a refreshed Siri interface, improved natural language processing, and AI-generated summaries, marks its foray into more advanced AI functionalities. Over the next year, Apple plans to introduce features driven by generative AI, such as image and emoji generation, and an enhanced Siri capable of accessing personal information and executing tasks within apps.

This is the second technical paper Apple has released on its AI systems, following a general overview published in June. The company is expected to provide more insights into its AI advancements when it reports its quarterly results later this week.

By leveraging Google’s TPUs, Apple demonstrates its commitment to exploring diverse and efficient solutions for AI training, setting the stage for further innovations in its AI-driven products and services.