Here’s the translation to American English:
Oracle has announced the general availability of its new Oracle Cloud Infrastructure (OCI) supercluster equipped with powerful Nvidia H200 GPUs. This supercluster, which can scale up to 65,536 H200 GPUs, promises unprecedented performance for artificial intelligence (AI) applications, reaching up to 260 exaflopsAn exaflop is a unit of measure in supercomputing that … of peak performance at FP8 precision, the company has reported.
The largest AI cloud infrastructure
Oracle claims that this supercluster is currently the largest cloud-based supercomputer for AI. Each compute instance within the supercluster delivers 76% more high-speed memory and 40% more bandwidthBandwidth is the maximum transfer capacity of… memory compared to H100 instances, improving inference performance in large language model (LLM) tasks by up to 1.9 times.
The system features a custom cluster network based on RDMA over Ethernet Converged Version 2 (RoCE v2), utilizing Nvidia ConnectX-7 network interface cards. This architecture enables interconnections between GPUs of up to 400 Gbps, while its front network of 200 Gbps facilitates the efficient transfer of large datasets between storage and GPUs.
Each bare metal instance is equipped with eight Nvidia H200 GPUs with 141 GB of HBM3e memory, along with two Intel Sapphire Rapids 8480+ CPUs with 56 cores.
Affordable costs and enhanced performance
Oracle maintains its competitive pricing policy: $10 per GPU per hour, the same cost as H100 instances. This offers companies more affordable access to cutting-edge AI infrastructure.
The supercluster also surpasses its predecessor, the H100, which could scale up to 16,384 GPUs, establishing itself as an ideal option for massive workloads such as training and inference of next-generation AI models.
Looking to the future: Nvidia Blackwell
In September 2024, Oracle revealed its plans to build an even more advanced supercluster, featuring up to 131,072 Nvidia Blackwell GPUs, scheduled for release in the first half of 2025. This development represents Oracle’s ongoing commitment to leading innovation in cloud computing for AI.
A leap toward the next generation of AI
Oracle’s supercluster with Nvidia H200 redefines the boundaries of cloud computing for artificial intelligence applications. With its scalable performance and competitive costs, it positions itself as a key tool for companies looking to leverage AI to tackle complex problems, from data analysis to generating advanced language models.
With this infrastructure, Oracle is not only addressing current demands for massive processing but is also laying the groundwork for future advancements in AI and high-performance computing.
via: DCD