Google A5X Instance Offers 1M Vera Rubin GPUs

Google announced the A5X Instance, a new infrastructure platform designed for large-scale artificial intelligence workloads. The company revealed the system supports up to 1,000,000 NVIDIA Vera Rubin GPUs for users. This architecture allows for massive computational capacity across distributed environments.

The platform features a single cluster that can accommodate up to 80,000 Rubin GPUs. It also supports multi-site clusters with access to up to 960,000 GPUs. The system utilizes NVIDIA ConnectX-9 network interface cards to manage data traffic. Google also confirmed the instance supports up to 134,000 TPUs within a single data center. Multi-site deployments can connect over 1,000,000 TPU chips. The hardware relies on a custom Arm-based CPU architecture. The system integrates an eighth-generation tensor processor for specialized tasks.

NVIDIA claims the infrastructure reduces AI inference costs per token by 10 times compared to the previous generation. The company also states the system improves throughput per megawatt by 10 times relative to older models. Pricing information for the A5X Instance is not currently available. Release dates for the A5X Instance have not been confirmed.

Google unveils A5X Instance supporting 1M Vera Rubin GPUs

The A5X Instance provides access to up to 1,000,000 NVIDIA Vera Rubin GPUs and over 1,000,000 TPU chips across multi-site clusters. The platform supports up to 80,000 Rubin GPUs per single cluster and over 1,000,000 chips across multi-site environments. The A5X Instance infrastructure supports up to 1,000,000 NVIDIA Vera Rubin GPUs, up to 80,000 Rubin GPUs per single cluster, and over 1,000,000 chips across multi-site environments.