Oracle and NVIDIA announced the largest cloud AI cluster to date, consisting of 131,072 NVIDIA B200 (Blackwell) accelerators. According to the companies, this is the world’s first 2.4 Zflops (FP8) system. The cluster will start operating in the first half of 2025, but the company is ready to accept orders for bare-metal instances and OCI Supercluster now. Customers will also be able to choose the connection type: RoCEv2 (ConnectX-7/8) or InfiniBand (Quantum-2).
According to the company, the new AI cluster is six times larger than what AWS, Microsoft Azure and Google Cloud can offer. In addition, the company offers other clusters with NVIDIA accelerators: 32768xA100, 16384xH100, 65536xH200 and 3840xL40S. And next year, clusters based on the GB200 NVL72 are promised, combining more than 100 thousand GB200 accelerators. Much more modest GPU.A100.1 and GPU.H100.1 VMs with one A100/H100 accelerator (80 GB) will also appear soon.
GPU.H200.8 instances are available for order right now, including eight H200 accelerators (141 GB), 30.7 TB of local NVMe storage and 200G connectivity. The family of instances based on NVIDIA Blackwell so far includes only two options. GPU.B200.8 offers eight B200 accelerators (192 GB), 30.7 TB of local NVMe storage and 400G connectivity. Finally, the GPU.GB200 is essentially the GB200 NVL72 super accelerator and includes 72 B200 accelerators, 36 Arm Grace processors and 533 TB of local NVMe storage. The aggregated network connection speed is 7.2 Tbit/s.
Image Source: Oracle
For all new instances, Oracle will provide managed Luster storage with performance up to 8 Gbps per TB. In addition, the company will offer advanced monitoring and management tools, assistance in setting up infrastructure to achieve the desired level of real-world performance, as well as a set of optimized software for working with AI, including for Arm.