Apple’s AI Journey: Leveraging Google Tensor Hardware
In a fascinating twist, Apple has been using Google Tensor hardware to lay the early foundations of its groundbreaking Apple Intelligence technology. Let’s delve into the details:
Research Paper Revelation:
- Apple recently published a research paper titled “Apple Intelligence Foundation Language Models.”
- Within this technical document, Apple provides insights into the sources of the language model at the core of its new AI technology.
Google Tensor Hardware Usage:
- Buried within the paper is a revealing quote: Apple’s Foundation Model (AFM) and the server technology driving it were initially constructed using Google’s Cloud TPUs.
- Specifically, Apple’s AFM and server components were built on v4 and v5p Cloud TPU clusters using Apple’s software.
Strategic Implications:
- While the initial training occurred on Google-designed hardware, Apple’s long-term strategy involves optimizing AI applications within its own data centers.
- Rumored as Project ACDC, Apple aims to match the technological capabilities of industry leaders like Microsoft and Meta by enhancing its AI server infrastructure.
- The company’s investment in AI server enhancements is set to exceed $5 billion over the next two years.
Data Compression and Acquisition:
- Apple has also acquired firms in Canada and France specializing in compressing data used in AI queries to data centers.
- By leveraging Google Tensor hardware during early development, Apple demonstrates a strategic move in the tech sector.
- In summary, Apple’s journey into AI involves a blend of collaboration and proprietary advancements. Whether it’s renting, acquiring, or optimizing hardware, the ultimate goal is to deliver cutting-edge AI experiences to users worldwide.
What is a Tenor (TPU) Chip?
A Tensor Processing Unit (TPU) is a custom-designed application-specific integrated circuit (ASIC) by Google that speeds up machine learning workloads. TPUs are ideal for accelerating AI training and inference because of their specialized features, such as the matrix multiply unit (MXU) and proprietary interconnect topology.
TPUs are designed to quickly perform matrix operations, which are fundamental in neural networks. They use a systolic array to perform large hard-wired matrix calculations without needing to access memory. The core of a TPU can perform up to 250,000 operations per clock cycle, even though its clock speed is only 700 megahertz.
TPUs can be used to run machine learning workloads using frameworks such as TensorFlow, Pytorch, and JAX. Google Cloud's Cloud TPU service makes TPUs available as scalable computing resources.
TPUs are considered to be AI's best friend and are lightning-fast for machine learning tasks. However, they are more expensive than CPUs and GPUs, though some say their pros outweigh their high price tag.