In a surprising move, Apple announced on Monday that it had chosen Google’s Tensor Processing Unit (TPU) for training the artificial intelligence models that power its Apple Intelligence system. This decision marks a significant shift away from the dominance of Nvidia’s GPUs in high-end AI training chips. The revelation came in a technical paper released by Apple, shedding light on the company’s strategy to explore alternatives to Nvidia for cutting-edge AI training.
Nvidia’s GPUs have long been the go-to choice for companies seeking high-performance AI training chips. However, the surge in demand for Nvidia’s GPUs over the past few years has made them increasingly challenging to acquire in sufficient quantities. Despite this, tech giants like OpenAI, Microsoft, and Anthropic continue to rely on Nvidia’s GPUs for their AI models.
Apple’s Embrace of Google’s TPUs
Apple’s decision to use Google’s TPUs for training its AI models comes as a surprise to many in the industry. While Apple did not explicitly mention Google or Nvidia in its technical paper, the company revealed that its Apple Foundation Model (AFM) and AFM server were trained on “Cloud TPU clusters.” This move allowed Apple to efficiently scale its training efforts, including on-device and server-based models.
Unlike its peers, Apple entered the generative AI space relatively later, following the launch of ChatGPT by OpenAI in late 2022. However, on Monday, Apple unveiled Apple Intelligence, introducing new features such as a revamped Siri interface, improved natural language processing, and AI-generated summaries in text fields. Over the next year, Apple plans to expand its generative AI capabilities to include image and emoji generation, as well as enhanced Siri functionalities.
Google’s TPUs have gained popularity for their cost efficiency, with the latest models costing under $2 per hour of use when booked for three years in advance. Initially introduced for internal workloads in 2015, Google made its TPUs available to the public in 2017. These custom chips are now considered among the most mature for artificial intelligence applications, highlighting Google’s commitment to advancing AI infrastructure.
Apple’s technical paper revealed that AFM on-device was trained on a single “slice” of 2048 TPU v5p chips, while AFM-server utilized 8192 TPU v4 chips configured into eight slices across a data center network. This comprehensive training approach demonstrates Apple’s dedication to leveraging advanced AI technologies to enhance its AI offerings.
Apple’s decision to adopt Google’s TPUs for training its AI models showcases the company’s willingness to explore alternatives to Nvidia in the AI training space. By embracing Google’s TPUs and expanding its generative AI capabilities, Apple is positioning itself for future success in the rapidly evolving AI industry.
Leave a Reply