Apple recently disclosed that it opted for Google TPUs rather than Nvidia hardware accelerators to develop its Apple Intelligence features. This was highlighted in an official research document, revealing the use of Google TPUv4 and TPUv5 chips in crafting the Apple Foundation Models (AFMs).
These models, including server and on-device versions, were essential for new online and offline intelligence features introduced at WWDC 2024. The largest model, the AFM-server, underwent training on 8,192 TPUv4 chips through a triple-stage pre-training process involving billions of tokens.
Training data comprised information from Applebot web crawler, other high-quality, licensed datasets, and select code and public datasets.
Read more – Apple skips Nvidia’s GPUs for its AI models, uses thousands of Google TPUs instead
