Apple Chooses Google TPUs Over Nvidia GPUs for AI Model Training

Apple's research reveals a strategic shift in AI hardware, utilizing Google's tensor processing units to develop its latest AI models.

Apple used 2,048 TPUv5p chips for on-device AI model training and 8,192 TPUv4 chips for server-side models.
The decision marks a departure from industry norms where Nvidia GPUs are typically favored for AI training.
Google's TPUs were accessed through its cloud platform, requiring software development within Google's ecosystem.
Apple's AI models, part of Intelligence, are being rolled out to beta users this week.
The research paper highlights potential for creating even larger models with Google's hardware.