Apple has revealed its decision to utilise Google’s Tensor Processing Units (TPUs) for training its AI models, eschewing the industry-standard NVIDIA GPUs. This revelation came to light through a technical paper published by Apple on Monday, detailing the company’s approach to developing its AI capabilities.
The Cupertino-based tech giant disclosed that it employed Google’s cloud-based TPU clusters, specifically the v4 and v5p versions, to train its Apple Foundation Model (AFM). “The AFM models are pre-trained on v4 and v5p Cloud TPU clusters with the AXLearn framework [Apple, 2023], a JAX [Bradbury et al., 2018] based deep learning library designed for the public cloud,” mentioned the company in the research paper.
This model underpins Apple’s new AI initiative, dubbed Apple Intelligence, which is set to power a range of upcoming AI-driven features across Apple’s product ecosystem.
Apple’s decision to use Google’s TPUs instead of NVIDIA’s GPUs is significant, given NVIDIA’s dominance in the AI chip market. NVIDIA’s GPUs are the top choice for AI training, powering much of the world’s leading AI research. The company controls about 95% of the AI chip market.
The decision suggests that major tech companies are exploring alternatives to NVIDIA’s offerings. It also highlights the growing capabilities of Google’s custom-designed TPUs, which were initially developed for internal use but are now available to external customers through Google Cloud Platform. Similarly, Meta has recently partnered with Groq, known for its LPUs, for running and inferring Llama 3.1.
Apple’s research paper indicates that for its server-based AI model, the company utilized 8,192 TPUv4 chips. For the on-device AI model designed for iPhones and other Apple devices, 2,048 TPUv5p chips were employed. These substantial deployments underscore the scale of Apple’s AI ambitions and the computational power required to train advanced language models.
Industry analysts speculate that Apple’s decision may be driven by factors such as cost-efficiency, performance advantages, or strategic partnerships. However, the company has not provided specific reasons for choosing Google’s TPUs over NVIDIA’s GPUs.
This development comes as Apple begins to roll out elements of its Apple Intelligence framework to beta users, signaling the company’s intent to compete more aggressively in the AI space.
As the AI arms race intensifies among tech giants, Apple’s hardware choices for its AI infrastructure will likely be closely watched by industry observers and competitors alike. The company’s decision to partner with Google for AI training hardware, while developing its own chips for on-device AI processing, demonstrates a nuanced approach to building out its AI capabilities.