Google Cloud and NVIDIA Launch Vera Rubin Infrastructure

Google Cloud and NVIDIA have announced the general availability of the A5X instances, which will be available next month, with an extension of availability to Europe. These integrated platforms are designed for training, inference and large-scale reinforcement learning. The A5X instances combine the NVIDIA Vera Rubin NVL72, NVIDIA ConnectX-9 SuperNICs, and Google Cloud's advanced liquid cooling technology, targeting demanding inference and reinforcement learning workloads.

Google Cloud plans to expand the Vera Rubin-based infrastructure to utilize up to 960,000 Rubin GPUs at various locations. The London-based company Ineffable Intelligence, founded by David Silver, the creator of AlphaGo, was the first customer to gain access to the A5X instances. Ineffable uses this platform to promote large-scale reinforcement learning and to develop 'superlearner' systems that continuously learn from experience.

The A5X instance supports a continuous cycle of learning through interaction, where experiences are generated, results are evaluated, and models are updated. This requires infrastructure that can support the entire process at scale. David Silver, founder and CEO of Ineffable, emphasizes that combining high computing capacity and high bandwidth connectivity with the A5X instance is necessary to advance the scale and ambition of their reinforcement learning research.

Another example of application is WPP, a marketing and communications company. They demonstrated their creative agent reference architecture on the A5X instances by utilizing NVIDIA Nemotron models and NVIDIA OpenShell. This technology enables conversion of conversation requests into complex creative workflows, with human approval for execution. Furthermore, Google Cloud and NVIDIA plan to offer NVIDIA Confidential Computing on bare metal instances and G4 VMs, allowing customers to benefit from confidential inference and fine-tuning.

Additionally, the two companies are exploring the possibilities of NVIDIA DSX MaxLPS for energy-efficient implementation of the Vera Rubin infrastructure, to optimize performance per watt and to expand capacity within available data center energy budgets.

Read the full article from Nvidia.