Apple’s technical paper, released on Monday, revealed that their two artificial intelligence models were pre-trained on Google’s cloud chips. One of these chips, called TPU (Tensor Processing Units), is designed by Google. Google’s TPU was launched in 2015 and publicly released in 2017. Initially created for internal workloads, it now has a wider range of applications. Apple’s move confirms that top technology companies in the field of AI training models are looking for alternatives to Nvidia’s GPU chips.
Nvidia’s expensive graphics processing units (GPU) dominate the high-end AI training chip market and have experienced high demand in recent years, resulting in a shortage of supply. OpenAI, Microsoft, and Anthropic all use Nvidia’s GPU in their models. Other tech companies, including Google, Meta, Oracle, and Tesla, also compete to purchase Nvidia chips to build their AI systems and products.
While Apple did not explicitly state that they did not use any Nvidia chips, the company was not mentioned in their description. Apple’s 47-page technical document did not mention Google or Nvidia, but it did indicate that their Apple Foundation Model (AFM) and AFM servers were trained on Google’s cloud TPU cluster. This means that Apple rented servers from a cloud provider to perform computations. Apple stated in the article that this system allows them to efficiently scale the training of AFM models, including AFM on devices, AFM servers, and larger models.
According to Reuters, Apple’s report indicates that they will rely on Google chips instead of Nvidia’s industry-leading graphic chips to build the infrastructure for their upcoming Apple Intelligence. This decision to rely on Google’s cloud infrastructure is noteworthy because Nvidia holds around 80% market share and produces the most popular AI graphics processors. Chips produced by Google, Amazon, and other cloud computing companies also require cooperation with Nvidia.
In their research report, Apple did not explicitly state that they did not use Nvidia chips, but their description of AI hardware and software did not mention Nvidia’s hardware devices. To train their AI models, the iPhone manufacturer used two forms of Google’s Tensor Processing Units (TPU), which were organized into large chip clusters. Apple used 2,048 TPUv5p chips to build AI models that can run on iPhones and other devices. They also deployed 8,192 TPUv4 processors for their server AI models.
Nvidia focuses on graphics processing units (GPU) widely used for AI tasks. Google and Nvidia differ in their sales approach, with Google selling access to TPUs through Google Cloud Platform. Customers interested in purchasing access to TPUs must establish software through Google’s cloud platform system to use these chips.
Apple has introduced partial Apple Intelligence features to its beta users. Reuters reported on the use of TPU chips in June, but Apple’s recently published research disclosed the extent of their dependence on Google’s hardware.
Apple engineers stated in the report that using Google’s chips allows them to create larger and more complex models than the two models studied in the report. Apple introduced a series of new AI features, including the integration of OpenAI’s ChatGPT technology into their software, at the June developer conference.
Investors are concerned about the future application and development of Apple Intelligence on iPhones. Apple introduced Apple Intelligence at the 2024 developer conference, integrating it into iOS 18, iPad OS 18, and macOS Sequoia as a personal intelligence system. Apple Intelligence consists of multiple generative models for users’ daily tasks and dynamically learns and releases their current tasks. The built-in foundation models fine-tune users’ actions, such as writing and polishing texts, prioritizing and summarizing notifications, creating interesting images for conversations with family and friends, and simplifying built-in app features to reduce the frequency of switching between apps.
Apple’s recent new iPhones and Siri features have been underwhelming. The newly introduced Apple Intelligence is only compatible with iOS 18 and requires iPhone 15 Pro or higher, as well as macOS 15 and iPadOS 17 with M1 or higher. Experts noted that Apple’s delay in launching Apple Intelligence until October means they will miss the timing of the upcoming iPhone firmware update in September. Developers can only start testing the new AI features this week at the earliest, and investors and Apple fans are hopeful that the new Apple iPhone 16 scheduled to be released at the end of the year will have the enhanced AI capabilities of Apple Intelligence, although they might end up disappointed.
Apple’s slow progress in hardware and software development, as well as their close relationship with Google, raises the question of whether Apple Intelligence will be a remarkable product that surpasses Siri or just a rehashed version of Siri 2.0. The iPhone, once a groundbreaking product introduced by Steve Jobs, is gradually losing its status as the only choice for smartphone users in the era of rapid AI development. Apple’s innovative AI features will undoubtedly impact their position and stock price in the tech industry.
Apple Intelligence, Google TPU Tensor Processing Units, artificial intelligence, Apple iPhone, Nvidia