No love lost between Apple and Nvidia

Apple has opted to use Google's Tensor Processing Units (TPUs) for training its generative AI models, as detailed in a recent research paper titled "Apple Intelligence Foundation Language Models." This decision marks a significant departure from the industry trend of utilizing Nvidia GPUs, particularly the H100, which are favored by many AI organizations. Apple's choice of Google's TPU v4 and TPU v5 processors comes amid ongoing tensions between Apple and Nvidia, which have persisted for several years. The paper reveals that Apple's server-side AI model, AFM-server, was trained on 8,192 TPU v4 chips, while the on-device model utilized 2,048 TPU v5 processors. Apple claims that its models often outperform those from competitors like Meta, OpenAI, and Google, although the research indicates mixed results in head-to-head comparisons. For instance, while the AFM-on-device model performed well against smaller models, it struggled against larger ones like GPT-4. Apple also highlighted its models' ability to generate safer content, reporting lower rates of harmful responses compared to other models. Despite the collaboration with Google for hardware, Apple recently criticized the company over user privacy issues related to online advertising. This move reflects Apple's strategic focus on developing its AI capabilities while navigating complex relationships within the tech industry.