Apple Uses Google's Tensor Chip To Develop Apple Intelligence

Apple revealed that it uses the Tensor Processing Unit (TPU) developed by Google, (photo; x @teknobeyin_)

JAKARTA - On July 30, 2024, at 5:56 AM PDT, Apple revealed that they used the Tensor Processing Unit (TPU) developed by Google, instead of a graphics processing unit (GPU) widely used from Nvidia, to build two important components of Apple Intelligence.

This decision is outlined in a new research paper published by Apple, which highlights its dependence on cloud hardware from Google. The paper reveals that Apple utilized 2,048 TPuv5p chips from Google to build AI models and 8,192 TPuv4 processors for server AI models. Although the paper does not explicitly mention Nvidia, the absence of reference to Nvidia hardware in Apple's AI infrastructure description shows a deliberate choice to prioritize technology from Google.

This decision is interesting given Nvidia's dominance in the AI processor market and because Apple rarely discloses its hardware choices for development purposes. The Nvidia GPU is in great demand for AI applications because of its high performance and efficiency.

Unlike Nvidia, which sells chips and their systems as independent products, Google provides access to its TPU through cloud services. Customers using Google TPU must develop their software in the Google ecosystem, which offers integrated tools and services to simplify the development and implementation of AI models.

In the paper, Apple engineers explained that TPU allows them to efficiently train large and sophisticated AI models. They illustrate how Google TPU is becoming a big cluster, which allows the processing power needed to train Apple's AI model.

Apple has announced plans to invest more than $5 billion in an AI server upgrade over the next two years, which should strengthen their AI capabilities and reduce dependence on external hardware providers.

In addition to detailing Google's use of TPU, the paper also discusses ethical considerations in AI development. Apple emphasizes its adherence to responsible data practices, claiming that no personal user data is used in their AI model training.

The company relies on a mix of publicly available data, licensed, and open-source for training purposes. Apple added that its training data set, which includes publicly available web data and licensed content, is carefully selected to protect user privacy.