“OCI Zettascale10 was designed with the objective of integrating large-scale generative AI use circumstances, together with coaching and working massive language fashions,” stated Information-Tech’s Palanichamy.
Oracle additionally launched new capabilities in Oracle Acceleron, its OCI networking stack, that it stated helps prospects run workloads extra shortly and cost-effectively. They embrace devoted community materials, converged NICs, and host-level zero-trust packet routing that Oracle says can double community and storage throughput whereas chopping latency and value.
Oracle’s zettascale supercomputer is constructed on the Acceleron RoCE (RDMA over Converged Ethernet) structure and Nvidia AI infrastructure. This enables it to ship what Oracle calls “breakthrough” scale, “extraordinarily low” GPU-to-GPU latency, and improved value/efficiency, cluster use, and general reliability.
The brand new structure has a “broad, shallow, resilient” material, in keeping with Oracle, and takes benefit of switching capabilities constructed into fashionable GPU community interface playing cards (NICs). This implies it might hook up with a number of switches on the identical time, however every swap stays by itself remoted community airplane.
Prospects can thus deploy bigger clusters, quicker, whereas working into fewer stalls and checkpoint restarts, as a result of site visitors might be shifted to totally different community planes and re-routed when the system encounters unstable or contested paths.
The structure additionally options power-efficient optics and is “hyper-optimized” for density, as its clusters are situated in massive information middle campuses inside a two-kilometer radius, Oracle stated.
