- Apple’s AI system, Apple Intelligence, used Google’s Tensor Processing Units (TPUs) for initial model training.
- The shift indicates a move away from Nvidia’s GPUs, which dominate the high-end AI training market.
- Apple’s technical paper reveals training on “Cloud TPU clusters” involving advanced TPU v5p chips and TPU v4 chips.
- New features in Apple Intelligence include updated Siri, improved natural language processing, and AI-generated text summaries.
- Google’s TPUs, available since 2017, are among the most advanced for AI, costing under $2 per hour with long-term bookings.
- Despite using TPUs, Google remains a significant Nvidia customer, utilizing Nvidia GPUs for AI training and cloud services.
Main AI News:
Apple has revealed that its artificial intelligence system, Apple Intelligence, utilized Google’s custom Tensor Processing Units (TPUs) for the initial training of its AI models. This decision highlights a shift in Big Tech’s approach to AI training, signaling a move away from Nvidia’s dominant GPUs.
The use of Google’s TPUs, as detailed in a newly published technical paper by Apple, underscores the growing trend of tech giants seeking alternatives to Nvidia’s GPUs. This shift comes in response to the high demand and limited availability of Nvidia’s high-end graphics processing units, which have been crucial for training sophisticated AI models.
While Nvidia GPUs remain the industry standard, with companies like OpenAI, Microsoft, and Anthropic relying on them, other tech giants, including Google, Meta, Oracle, and Tesla, are also investing heavily in AI infrastructure. Last week, both Meta’s Mark Zuckerberg and Alphabet’s Sundar Pichai expressed concerns over potential overinvestment in AI infrastructure, despite recognizing the high stakes of falling behind.
Apple’s technical paper, although not naming Google or Nvidia explicitly, reveals that the company’s Apple Foundation Model (AFM) and AFM server were trained on “Cloud TPU clusters.” This setup involved renting servers to carry out the extensive calculations required. Apple’s approach, focusing on efficient and scalable training, includes models trained on the latest TPU v5p chips and TPUs configured in data center networks.
In its strategic move, Apple has also introduced a preview version of Apple Intelligence, featuring enhanced natural language processing, updated Siri functionalities, and AI-generated text summaries. Over the coming year, Apple plans to expand its AI capabilities with features like image and emoji generation, and an upgraded Siri capable of interacting with personal information and app functionalities.
Google’s TPUs, available since 2015 and accessible to the public since 2017, are now among the most advanced custom chips for AI, costing under $2 per hour when booked long-term. Despite its investment in TPUs, Google remains a significant Nvidia customer, using GPUs for its AI training and offering Nvidia’s technology through its cloud services.
Apple’s move to utilize Google’s TPUs reflects a broader industry trend towards diversifying AI infrastructure, a strategy that may reshape how tech companies approach AI development and deployment in the future.
Conclusion:
Apple’s choice to use Google’s TPUs for training its AI models represents a significant shift in the AI infrastructure landscape. This move highlights a growing trend among tech giants to explore alternatives to Nvidia’s GPUs, driven by the high demand and limited availability of Nvidia’s hardware. As companies like Apple and Google invest in diverse AI technologies, the market may see increased competition and innovation, potentially reshaping how AI systems are developed and deployed. This diversification could lead to more options for AI infrastructure, fostering a more competitive environment and accelerating advancements in AI technology.
