Apple skips Nvidia's GPUs for its AI models, uses thousands of Google TPUs instead
Apple has disclosed that its Apple Intelligence features were developed using Google TPUs rather than Nvidia hardware accelerators. This strategic choice highlights a significant shift in infrastructure reliance for major tech firms, demonstrating that high-performance AI training can be achieved through alternative specialized processors. The revelation underscores the diversity of available computational resources in the current AI landscape. The development process involved distinct clusters for server-side and on-device models. Massive TPUv4 arrays trained the foundational language models for online services, while optimized TPUv5 clusters facilitated the creation of efficient, smaller models for local execution. This dual approach ensures robust performance across both cloud-based and private, offline contexts. This transparency is highly relevant to open data discussions as it exposes the specific data sources, including web crawlers and licensed datasets, used to train these systems. By revealing the technical architecture and data provenance, Apple provides a case study on how large-scale proprietary AI interacts with public information ecosystems. It invites scrutiny regarding data quality and sourcing practices within the broader open data community.
Source: tomshardware.comPublished on 2024-07-31