LAS VEGAS – Oracle on Tuesday unveiled OCI Zettascale10, describing it as the largest artificial intelligence supercomputer in the cloud. The system connects hundreds of thousands of NVIDIA graphics processing units, or GPUs, to deliver up to 16 zettaFLOPS of peak performance, the company said.
Oracle said OCI Zettascale10 serves as the computing foundation for its flagship supercluster built with OpenAI in Abilene, Texas, under a project called Stargate. The system is based on Oracle’s Acceleron RoCE networking architecture and NVIDIA’s AI infrastructure.
Mahesh Thiagarajan, executive vice president for Oracle Cloud Infrastructure, said the new platform combines scale and efficiency for large AI workloads. “With OCI Zettascale10, we’re fusing Oracle’s Acceleron RoCE network architecture with next-generation NVIDIA AI infrastructure to deliver multi-gigawatt AI capacity at unmatched scale,” he said.
Large-scale AI infrastructure
Oracle said Zettascale10 connects GPUs across data centres within a two-kilometre radius to improve latency and power efficiency. The clusters are housed in gigawatt-scale data centre campuses optimised for AI training and inference workloads.
The company said customers will be able to build, train and deploy large AI models with lower power consumption per unit of performance. Oracle also said the system supports data sovereignty and security controls across its distributed cloud.
OpenAI’s vice president of infrastructure and industrial compute, Peter Hoeschele, said the Zettascale10 network fabric was first deployed at the Stargate site. He said the design “maximises fabric-wide performance at gigawatt scale while keeping most of the power focused on compute.”
NVIDIA partnership and deployment
Oracle said initial deployments of OCI Zettascale10 will feature up to 800,000 NVIDIA GPUs. The company said the architecture delivers high bandwidth between GPUs and predictable performance through Oracle Acceleron’s low-latency RoCEv2 networking.
Ian Buck, vice president of hyperscale at NVIDIA, said the collaboration would combine Oracle’s distributed cloud with NVIDIA’s full-stack AI infrastructure. “OCI Zettascale10 provides the compute fabric needed to advance state-of-the-art AI research and help organisations move from experimentation to industrialised AI,” he said.
Oracle said it plans to offer multi-gigawatt deployments of Zettascale10 to enterprise customers and is now taking orders for delivery in the second half of next year.
Networking technology
Oracle said its Acceleron RoCE networking architecture allows GPU network interface cards to connect to multiple switches across isolated network planes, increasing reliability and reducing downtime. The system can redirect traffic when one plane experiences problems, avoiding restarts during large AI training runs.
Other features include power-efficient optical connections, simplified maintenance through plane-level updates and a design that reduces network tiers to cut cost and improve consistency in GPU-to-GPU communication.
Oracle said the technology allows customers to deploy large AI clusters faster and operate them more efficiently by keeping more of the power budget focused on computation rather than networking.
