Google Unveils TPU 8t, 8i Chips with 9,600-Unit Scalability and Bandwidth Gains
Alphabet introduced two new TPU models, the 8t for large-model training and the 8i for real-time inference, with the 8t scaling up to 9,600 chips per system and the 8i boosting memory bandwidth for real-time tasks. The move aligns with Google's strategy to in-source hardware stack and upgrade networking.
1. TPU Chip Launch
Alphabet introduced TPU 8t for large-model training and TPU 8i for inference, featuring scalability up to 9,600 chips in a single system and enhanced memory bandwidth for real-time tasks.
2. Performance and Use Cases
The TPU 8t targets AI model training workloads while the 8i boosts latency-sensitive inference, addressing rising demand for computing power and positioning Google to optimize costs versus external suppliers.
3. Strategic Hardware and Network Upgrade
These chip launches align with Google's push to internalize its hardware stack and roll out upgraded data-center networking, aiming to reduce reliance on third-party providers and improve performance across its global infrastructure.