In a strategic move that underscores the shifting dynamics within the tech industry, Apple announced on Monday that its artificial intelligence models, which power the newly unveiled Apple Intelligence system, were pretrained on Google's Tensor Processing Units (TPUs). This development marks a significant pivot as major tech companies seek alternatives to Nvidia's dominant but costly GPUs for AI training.
The revelation came through a technical paper from Apple, detailing the use of Google's homegrown TPUs for training its AI models. The paper highlights Apple's decision to rent servers from a cloud provider to facilitate the training process. This choice is particularly noteworthy given the intense competition and high demand for Nvidia's GPUs, which have become a cornerstone for AI training across the industry.
Nvidia's GPUs have long been the go-to hardware for AI training, thanks to their unparalleled performance. Companies like OpenAI, Microsoft, and Anthropic have heavily relied on Nvidia's technology to develop their AI models. However, the high costs and limited availability of these GPUs have prompted some tech giants to explore other options. Google, Meta, Oracle, and Tesla have all been snapping up Nvidia's GPUs to bolster their AI capabilities, but Apple's latest move indicates a growing interest in diversifying AI training hardware.
The CEOs of Meta and Alphabet recently expressed concerns about potential overinvestment in AI infrastructure. Nevertheless, they acknowledged the critical importance of staying ahead in the AI race. Falling behind, they warned, could leave companies out of position for what is poised to be the most transformative technology of the next decade and beyond.
Apple's technical paper, while not explicitly naming Google or Nvidia, reveals that its Apple Foundation Model (AFM) and AFM server were trained on "Cloud TPU clusters." This setup allowed Apple to efficiently scale its AI models, including the AFM-on-device and AFM-server variants. The paper also detailed the use of a single "slice" of 2048 TPU v5p chips for training AFM on-device, and 8192 TPU v4 chips configured as eight slices for AFM-server training.
The decision to use Google's TPUs is a strategic one. Google's TPUs, introduced in 2015 for internal workloads and made publicly available in 2017, are among the most mature custom chips designed for AI. They offer a cost-effective alternative to Nvidia's GPUs, with the latest TPUs costing under $2 per hour when booked for three years in advance.
Despite its investment in TPUs, Google remains one of Nvidia's top customers, using both Nvidia's GPUs and its own TPUs for AI training. Google also offers access to Nvidia's technology through its cloud services, highlighting the complex and intertwined nature of the AI hardware market.
Apple's announcement of Apple Intelligence on Monday included a preview version for some devices, showcasing several new features. These include a refreshed look for Siri, improved natural language processing, and AI-generated summaries in text fields. Over the next year, Apple plans to roll out additional generative AI functions, such as image and emoji generation, and a more powerful Siri capable of accessing personal information and performing actions within apps.
This is the second technical paper Apple has published about its AI system, following a more general version released in June. At that time, Apple indicated its use of TPUs for AI model development. The company has also stated that inferencing—running a pretrained AI model to generate content or make predictions—will partially occur on Apple's own chips in its data centers.
Apple's decision to embrace TPUs for AI training reflects a broader trend in the tech industry. As AI continues to evolve and expand, companies are seeking more efficient and scalable solutions to meet their growing demands. The use of TPUs allows Apple to train its models more cost-effectively and with greater flexibility, positioning the company to compete in the rapidly advancing AI landscape.
The announcement comes at a pivotal moment for Apple, which has been relatively late to the generative AI party compared to its peers. The launch of Apple Intelligence signals the company's commitment to integrating AI more deeply into its ecosystem. With new features and capabilities on the horizon, Apple is poised to make significant strides in the AI arena.
As the tech industry continues to navigate the complexities of AI development, Apple's strategic choice to leverage Google's TPUs highlights the importance of innovation and adaptability. By exploring alternative hardware solutions, Apple is not only addressing the challenges posed by Nvidia's GPU dominance but also paving the way for more diverse and competitive AI advancements.
With quarterly results set to be reported after the close of trading on Thursday, all eyes will be on Apple to see how its latest AI initiatives impact its financial performance and market position. The company's foray into AI, backed by the power of Google's TPUs, marks a new chapter in its ongoing quest to push the boundaries of technology and enhance the user experience.