OCI redefines top-level AI performance with Zettascale10 and Acceleron

LAS VEGAS—Get ready for massive amounts of AI-enabling cloud computing power. Oracle announced its OCI Zettascale10 supercluster, which delivers 16 zettaFLOPS within a multigigawatt cluster, at its AI World conference this week. The hyperscaler also launched a high-speed Ethernet-based network interconnect, called Oracle Acceleron, as well as a new multicloud software licensing system that offers customers an unprecedented range of deployment options.

The Zettacale10 Supercluster, built on Oracle Cloud Infrastructure (OCI), spans multiple data centers and integrates as many as 800,000 NVIDIA GPUs, all tied together using Oracle’s ultra-low-latency InfiniBand and RoCE-based networks. InfiniBand is an industry-standard high-speed interconnect for supercomputers that lets one server read or write another server’s memory directly using RDMA, or Remote Direct Memory Access. Oracle Acceleron RoCE (RDMA over Converged Ethernet) networking extends RDMA into cloud computing clusters and is particularly well-suited to AI tasks.

“Customers can build, train, and deploy their largest AI models into production using less power per unit of performance and achieving higher reliability,” said Mahesh Thiagarajan, executive vice president, OCI. “With OCI Zettascale10, we’re fusing OCI’s groundbreaking Oracle Acceleron RoCE network architecture with next-generation NVIDIA AI infrastructure to deliver multigigawatt AI capacity at unmatched scale.”

The power of Zettascale10 isn’t measured merely in zettaFLOPS (floating point operations per section) or even gigawatts. A Zettascale10 cluster lives in many densely packed data centers, all interconnected with high-capacity fiber optic links. The upshot: a supercluster that can handle even the most challenging AI workloads.

The first Zettascale10 supercomputer is based in Abilene, Texas, in a large gigawatt data center campus.

Acceleron RoCE scales the compute fabric

Key to the Zettascale10 supercluster is the latest in high-speed networking interconnects, Oracle Acceleron RoCE network fabric. Unlike InfiniBand, which offers excellent performance but requires specialized hardware, Acceleron uses specialized Ethernet network interface cards (NICs) and cables to link high performance GPUs within an OCI data center. Those specialized NICs include their own four-port Ethernet switches, a design that lowers the latency of traffic across a large, complex network.

With the Acceleron network fabric, Oracle can help customers deploy very large AI clusters quickly, and at lower cost, by connecting GPU NICs to multiple network switches simultaneously, thereby creating a flatter network fabric with many isolated network planes. That architecture lets traffic flow more rapidly, so GPUs aren’t stalled waiting to send or receive data. One benefit is efficiency—the Zettascale10 supercluster can deliver results significantly faster. Another is predictability: Performance is more uniform across supercomputer clusters, with lower GPU-to-GPU latency because there are fewer switches to traverse. This also adds resiliency in case part of the network stalls due to heavy utilization. Finally, the Acceleron RoCE fabric keeps data flowing even if one network plane is taken out of service for hardware or software maintenance.

Acceleron’s unique architecture delivers line-rate encryption through and across the network, with ultra-low latency. Customers get twice the potential networking capacity, up to twice the storage IOPS (input/output per second), and zero-trust packet routing (ZPR) security policies enforced directly at the host—that is, at every Acceleron NIC and every point on the network. One benefit of this approach: On-NIC security policies can block object storage access across the internet to prevent data exfiltration.

“Acceleron is the foundation for all our I/O security and acceleration functions,” said Oracle CEO Clay Magouyrk during his keynote at AI World. “Customers will see higher peak performance with lower cost, better ease of use, and increased security.”

Universal way to buy

Complexity defines cloud deployments for many customers that use a diverse web of applications and services from hyperscalers Oracle, Amazon Web Services, Google Cloud, and Microsoft Azure. Even when an architecture must be complicated to meet an organization’s technical requirements, service licensing should be simple, and that’s where the new Oracle Multicloud Universal Credits program helps.

With the program, customers procure Oracle Database and OCI services in the cloud, or clouds, of their choice. The credits are usable on Oracle Database@AWS, Oracle Database@Azure, Oracle Database@Google Cloud, and OCI itself. That lets customers flexibly deploy workloads and see consistent contracts across clouds.

“With 37 multicloud regions now live across AWS, Azure, and Google Cloud, and the coming launch of Oracle Multicloud Universal Credits, we’re giving customers more choices and flexibility than ever by simplifying contracts and introducing the industry’s first flexible, cross-cloud consumption model,” said Karan Batta, senior vice president, OCI. With a single licensing and consumption model across clouds, customers gain a single administration and governance structure for Oracle Applications and services, as well as a consistent database experience.

Companies can also use their credits for any OCI service in the hyperscaler cloud, or clouds, of their choice.

New performance, flexibility standards for enterprise AI

With the Zettascale10 supercomputer and Acceleron RoCE network fabric technology, customers can run heavy AI workloads across as many as 800,000 GPUs. With the Oracle Multicloud Universal Credits program, the licensing and use of Oracle Database services across multiple clouds has never been easier or more straightforward. With these advances, Oracle delivers efficiency, simplicity, flexibility, exceptional performance, and choice for multicloud customers.

View more Oracle Connect articles