Apple has released documentation detailing the development of their artificial intelligence models known as Apple Intelligence Foundation Language Models (AFM). A notable aspect is the technology Apple chose to use for training the AI models, as they did not use NVIDIA’s GPU. The documents indicate that Apple opted for the use of TPU chips developed by Google for model training. There are two main models: the on-device AFM trained with TPUv5p chips totaling 2,048 units, and the server-side AFM using TPUv4 with 8,192 units. Apple did not disclose the name of Google, as well as the acquisition method of these TPUs for model running, which Google only allows through their cloud platform. Apple did not explain the reason for choosing TPU over GPU for model training, but stated that utilizing TPUs provides efficient and scalable model training, especially for very large models that Apple aims for.
Source: CNBC
TLDR: Apple utilizes Google-developed TPUs for training their AI models instead of NVIDIA’s GPUs, citing efficiency and scalability benefits for large models.
Leave a Comment