Google AI 인프라의 미래: 에이전틱 시대를 위한 확장

Google announces AI Hypercomputer, an integrated infrastructure stack optimized for the agentic era with next-generation TPUs, GPUs, networking, and storage systems.

• •TPU 8t: Eighth-generation processor with 3x higher compute power (121 exaflops) supporting 100+ million TPU chip coordination for complex model training
• •TPU 8i: Inference and reinforcement learning processor with 80% improved cost efficiency, ultra-low latency, and dedicated acceleration engine for concurrent agentic workloads
• •Virgo Network: Datacenter fabric with 4x bandwidth enabling 1+ million TPU clustering across multiple sites for near-linear scaling
• •Enhanced storage: 10TB/s Managed Lustre, TPUDirect for direct GPU/TPU access, Rapid Buckets for sub-millisecond latency checkpoint operations
• •