Apple's Foundation Language Models Trained on Google's TPUs, Marking Shift from Nvidia GPUs

Jul 29, 2024, 09:19 PM

Apple has disclosed that its AI models, which power the Apple Intelligence features, were trained using Google's Tensor Processing Units (TPUs). The technical paper released by Apple details the use of Google's TPUs for both on-device and server versions of the models. Specifically, the on-device models were trained using 2048 TPU v5p chips, while the server models utilized 8192 TPU v4 chips. This marks a significant shift from the industry-standard Nvidia GPUs, which were not used in any part of the training process. The move highlights a growing trend among major tech companies to explore alternatives to Nvidia for AI training. Apple's Foundation Language Models were developed with a responsible approach to AI training.